Commit | Line | Data |
---|---|---|
c1d7c514 | 1 | // SPDX-License-Identifier: GPL-2.0 |
53b381b3 DW |
2 | /* |
3 | * Copyright (C) 2012 Fusion-io All rights reserved. | |
4 | * Copyright (C) 2012 Intel Corp. All rights reserved. | |
53b381b3 | 5 | */ |
c1d7c514 | 6 | |
53b381b3 | 7 | #include <linux/sched.h> |
53b381b3 DW |
8 | #include <linux/bio.h> |
9 | #include <linux/slab.h> | |
53b381b3 | 10 | #include <linux/blkdev.h> |
53b381b3 DW |
11 | #include <linux/raid/pq.h> |
12 | #include <linux/hash.h> | |
13 | #include <linux/list_sort.h> | |
14 | #include <linux/raid/xor.h> | |
818e010b | 15 | #include <linux/mm.h> |
9b569ea0 | 16 | #include "messages.h" |
cea62800 | 17 | #include "misc.h" |
53b381b3 | 18 | #include "ctree.h" |
53b381b3 | 19 | #include "disk-io.h" |
53b381b3 DW |
20 | #include "volumes.h" |
21 | #include "raid56.h" | |
22 | #include "async-thread.h" | |
c5a41562 | 23 | #include "file-item.h" |
7a315072 | 24 | #include "btrfs_inode.h" |
53b381b3 DW |
25 | |
26 | /* set when additional merges to this rbio are not allowed */ | |
27 | #define RBIO_RMW_LOCKED_BIT 1 | |
28 | ||
4ae10b3a CM |
29 | /* |
30 | * set when this rbio is sitting in the hash, but it is just a cache | |
31 | * of past RMW | |
32 | */ | |
33 | #define RBIO_CACHE_BIT 2 | |
34 | ||
35 | /* | |
36 | * set when it is safe to trust the stripe_pages for caching | |
37 | */ | |
38 | #define RBIO_CACHE_READY_BIT 3 | |
39 | ||
4ae10b3a CM |
40 | #define RBIO_CACHE_SIZE 1024 |
41 | ||
8a953348 DS |
42 | #define BTRFS_STRIPE_HASH_TABLE_BITS 11 |
43 | ||
44 | /* Used by the raid56 code to lock stripes for read/modify/write */ | |
45 | struct btrfs_stripe_hash { | |
46 | struct list_head hash_list; | |
47 | spinlock_t lock; | |
48 | }; | |
49 | ||
50 | /* Used by the raid56 code to lock stripes for read/modify/write */ | |
51 | struct btrfs_stripe_hash_table { | |
52 | struct list_head stripe_cache; | |
53 | spinlock_t cache_lock; | |
54 | int cache_size; | |
55 | struct btrfs_stripe_hash table[]; | |
56 | }; | |
57 | ||
eb357060 QW |
58 | /* |
59 | * A bvec like structure to present a sector inside a page. | |
60 | * | |
61 | * Unlike bvec we don't need bvlen, as it's fixed to sectorsize. | |
62 | */ | |
63 | struct sector_ptr { | |
64 | struct page *page; | |
00425dd9 QW |
65 | unsigned int pgoff:24; |
66 | unsigned int uptodate:8; | |
eb357060 QW |
67 | }; |
68 | ||
93723095 QW |
69 | static void rmw_rbio_work(struct work_struct *work); |
70 | static void rmw_rbio_work_locked(struct work_struct *work); | |
53b381b3 DW |
71 | static void index_rbio_pages(struct btrfs_raid_bio *rbio); |
72 | static int alloc_rbio_pages(struct btrfs_raid_bio *rbio); | |
73 | ||
6bfd0133 QW |
74 | static int finish_parity_scrub(struct btrfs_raid_bio *rbio, int need_check); |
75 | static void scrub_rbio_work_locked(struct work_struct *work); | |
5a6ac9ea | 76 | |
797d74b7 QW |
77 | static void free_raid_bio_pointers(struct btrfs_raid_bio *rbio) |
78 | { | |
2942a50d | 79 | bitmap_free(rbio->error_bitmap); |
797d74b7 QW |
80 | kfree(rbio->stripe_pages); |
81 | kfree(rbio->bio_sectors); | |
82 | kfree(rbio->stripe_sectors); | |
83 | kfree(rbio->finish_pointers); | |
84 | } | |
85 | ||
ff2b64a2 QW |
86 | static void free_raid_bio(struct btrfs_raid_bio *rbio) |
87 | { | |
88 | int i; | |
89 | ||
90 | if (!refcount_dec_and_test(&rbio->refs)) | |
91 | return; | |
92 | ||
93 | WARN_ON(!list_empty(&rbio->stripe_cache)); | |
94 | WARN_ON(!list_empty(&rbio->hash_list)); | |
95 | WARN_ON(!bio_list_empty(&rbio->bio_list)); | |
96 | ||
97 | for (i = 0; i < rbio->nr_pages; i++) { | |
98 | if (rbio->stripe_pages[i]) { | |
99 | __free_page(rbio->stripe_pages[i]); | |
100 | rbio->stripe_pages[i] = NULL; | |
101 | } | |
102 | } | |
103 | ||
104 | btrfs_put_bioc(rbio->bioc); | |
797d74b7 | 105 | free_raid_bio_pointers(rbio); |
ff2b64a2 QW |
106 | kfree(rbio); |
107 | } | |
108 | ||
385de0ef | 109 | static void start_async_work(struct btrfs_raid_bio *rbio, work_func_t work_func) |
ac638859 | 110 | { |
385de0ef CH |
111 | INIT_WORK(&rbio->work, work_func); |
112 | queue_work(rbio->bioc->fs_info->rmw_workers, &rbio->work); | |
ac638859 DS |
113 | } |
114 | ||
53b381b3 DW |
115 | /* |
116 | * the stripe hash table is used for locking, and to collect | |
117 | * bios in hopes of making a full stripe | |
118 | */ | |
119 | int btrfs_alloc_stripe_hash_table(struct btrfs_fs_info *info) | |
120 | { | |
121 | struct btrfs_stripe_hash_table *table; | |
122 | struct btrfs_stripe_hash_table *x; | |
123 | struct btrfs_stripe_hash *cur; | |
124 | struct btrfs_stripe_hash *h; | |
125 | int num_entries = 1 << BTRFS_STRIPE_HASH_TABLE_BITS; | |
126 | int i; | |
127 | ||
128 | if (info->stripe_hash_table) | |
129 | return 0; | |
130 | ||
83c8266a DS |
131 | /* |
132 | * The table is large, starting with order 4 and can go as high as | |
133 | * order 7 in case lock debugging is turned on. | |
134 | * | |
135 | * Try harder to allocate and fallback to vmalloc to lower the chance | |
136 | * of a failing mount. | |
137 | */ | |
ee787f95 | 138 | table = kvzalloc(struct_size(table, table, num_entries), GFP_KERNEL); |
818e010b DS |
139 | if (!table) |
140 | return -ENOMEM; | |
53b381b3 | 141 | |
4ae10b3a CM |
142 | spin_lock_init(&table->cache_lock); |
143 | INIT_LIST_HEAD(&table->stripe_cache); | |
144 | ||
53b381b3 DW |
145 | h = table->table; |
146 | ||
147 | for (i = 0; i < num_entries; i++) { | |
148 | cur = h + i; | |
149 | INIT_LIST_HEAD(&cur->hash_list); | |
150 | spin_lock_init(&cur->lock); | |
53b381b3 DW |
151 | } |
152 | ||
153 | x = cmpxchg(&info->stripe_hash_table, NULL, table); | |
fe3b7bb0 | 154 | kvfree(x); |
53b381b3 DW |
155 | return 0; |
156 | } | |
157 | ||
4ae10b3a CM |
158 | /* |
159 | * caching an rbio means to copy anything from the | |
ac26df8b | 160 | * bio_sectors array into the stripe_pages array. We |
4ae10b3a CM |
161 | * use the page uptodate bit in the stripe cache array |
162 | * to indicate if it has valid data | |
163 | * | |
164 | * once the caching is done, we set the cache ready | |
165 | * bit. | |
166 | */ | |
167 | static void cache_rbio_pages(struct btrfs_raid_bio *rbio) | |
168 | { | |
169 | int i; | |
4ae10b3a CM |
170 | int ret; |
171 | ||
172 | ret = alloc_rbio_pages(rbio); | |
173 | if (ret) | |
174 | return; | |
175 | ||
00425dd9 QW |
176 | for (i = 0; i < rbio->nr_sectors; i++) { |
177 | /* Some range not covered by bio (partial write), skip it */ | |
88074c8b QW |
178 | if (!rbio->bio_sectors[i].page) { |
179 | /* | |
180 | * Even if the sector is not covered by bio, if it is | |
181 | * a data sector it should still be uptodate as it is | |
182 | * read from disk. | |
183 | */ | |
184 | if (i < rbio->nr_data * rbio->stripe_nsectors) | |
185 | ASSERT(rbio->stripe_sectors[i].uptodate); | |
00425dd9 | 186 | continue; |
88074c8b | 187 | } |
00425dd9 QW |
188 | |
189 | ASSERT(rbio->stripe_sectors[i].page); | |
190 | memcpy_page(rbio->stripe_sectors[i].page, | |
191 | rbio->stripe_sectors[i].pgoff, | |
192 | rbio->bio_sectors[i].page, | |
193 | rbio->bio_sectors[i].pgoff, | |
194 | rbio->bioc->fs_info->sectorsize); | |
195 | rbio->stripe_sectors[i].uptodate = 1; | |
196 | } | |
4ae10b3a CM |
197 | set_bit(RBIO_CACHE_READY_BIT, &rbio->flags); |
198 | } | |
199 | ||
53b381b3 DW |
200 | /* |
201 | * we hash on the first logical address of the stripe | |
202 | */ | |
203 | static int rbio_bucket(struct btrfs_raid_bio *rbio) | |
204 | { | |
4c664611 | 205 | u64 num = rbio->bioc->raid_map[0]; |
53b381b3 DW |
206 | |
207 | /* | |
208 | * we shift down quite a bit. We're using byte | |
209 | * addressing, and most of the lower bits are zeros. | |
210 | * This tends to upset hash_64, and it consistently | |
211 | * returns just one or two different values. | |
212 | * | |
213 | * shifting off the lower bits fixes things. | |
214 | */ | |
215 | return hash_64(num >> 16, BTRFS_STRIPE_HASH_TABLE_BITS); | |
216 | } | |
217 | ||
d4e28d9b QW |
218 | static bool full_page_sectors_uptodate(struct btrfs_raid_bio *rbio, |
219 | unsigned int page_nr) | |
220 | { | |
221 | const u32 sectorsize = rbio->bioc->fs_info->sectorsize; | |
222 | const u32 sectors_per_page = PAGE_SIZE / sectorsize; | |
223 | int i; | |
224 | ||
225 | ASSERT(page_nr < rbio->nr_pages); | |
226 | ||
227 | for (i = sectors_per_page * page_nr; | |
228 | i < sectors_per_page * page_nr + sectors_per_page; | |
229 | i++) { | |
230 | if (!rbio->stripe_sectors[i].uptodate) | |
231 | return false; | |
232 | } | |
233 | return true; | |
234 | } | |
235 | ||
eb357060 QW |
236 | /* |
237 | * Update the stripe_sectors[] array to use correct page and pgoff | |
238 | * | |
239 | * Should be called every time any page pointer in stripes_pages[] got modified. | |
240 | */ | |
241 | static void index_stripe_sectors(struct btrfs_raid_bio *rbio) | |
242 | { | |
243 | const u32 sectorsize = rbio->bioc->fs_info->sectorsize; | |
244 | u32 offset; | |
245 | int i; | |
246 | ||
247 | for (i = 0, offset = 0; i < rbio->nr_sectors; i++, offset += sectorsize) { | |
248 | int page_index = offset >> PAGE_SHIFT; | |
249 | ||
250 | ASSERT(page_index < rbio->nr_pages); | |
251 | rbio->stripe_sectors[i].page = rbio->stripe_pages[page_index]; | |
252 | rbio->stripe_sectors[i].pgoff = offset_in_page(offset); | |
253 | } | |
254 | } | |
255 | ||
4d100466 QW |
256 | static void steal_rbio_page(struct btrfs_raid_bio *src, |
257 | struct btrfs_raid_bio *dest, int page_nr) | |
258 | { | |
259 | const u32 sectorsize = src->bioc->fs_info->sectorsize; | |
260 | const u32 sectors_per_page = PAGE_SIZE / sectorsize; | |
261 | int i; | |
262 | ||
263 | if (dest->stripe_pages[page_nr]) | |
264 | __free_page(dest->stripe_pages[page_nr]); | |
265 | dest->stripe_pages[page_nr] = src->stripe_pages[page_nr]; | |
266 | src->stripe_pages[page_nr] = NULL; | |
267 | ||
268 | /* Also update the sector->uptodate bits. */ | |
269 | for (i = sectors_per_page * page_nr; | |
270 | i < sectors_per_page * page_nr + sectors_per_page; i++) | |
271 | dest->stripe_sectors[i].uptodate = true; | |
272 | } | |
273 | ||
88074c8b QW |
274 | static bool is_data_stripe_page(struct btrfs_raid_bio *rbio, int page_nr) |
275 | { | |
276 | const int sector_nr = (page_nr << PAGE_SHIFT) >> | |
277 | rbio->bioc->fs_info->sectorsize_bits; | |
278 | ||
279 | /* | |
280 | * We have ensured PAGE_SIZE is aligned with sectorsize, thus | |
281 | * we won't have a page which is half data half parity. | |
282 | * | |
283 | * Thus if the first sector of the page belongs to data stripes, then | |
284 | * the full page belongs to data stripes. | |
285 | */ | |
286 | return (sector_nr < rbio->nr_data * rbio->stripe_nsectors); | |
287 | } | |
288 | ||
4ae10b3a | 289 | /* |
d4e28d9b QW |
290 | * Stealing an rbio means taking all the uptodate pages from the stripe array |
291 | * in the source rbio and putting them into the destination rbio. | |
292 | * | |
293 | * This will also update the involved stripe_sectors[] which are referring to | |
294 | * the old pages. | |
4ae10b3a CM |
295 | */ |
296 | static void steal_rbio(struct btrfs_raid_bio *src, struct btrfs_raid_bio *dest) | |
297 | { | |
298 | int i; | |
4ae10b3a CM |
299 | |
300 | if (!test_bit(RBIO_CACHE_READY_BIT, &src->flags)) | |
301 | return; | |
302 | ||
303 | for (i = 0; i < dest->nr_pages; i++) { | |
88074c8b QW |
304 | struct page *p = src->stripe_pages[i]; |
305 | ||
306 | /* | |
307 | * We don't need to steal P/Q pages as they will always be | |
308 | * regenerated for RMW or full write anyway. | |
309 | */ | |
310 | if (!is_data_stripe_page(src, i)) | |
4ae10b3a | 311 | continue; |
4ae10b3a | 312 | |
88074c8b QW |
313 | /* |
314 | * If @src already has RBIO_CACHE_READY_BIT, it should have | |
315 | * all data stripe pages present and uptodate. | |
316 | */ | |
317 | ASSERT(p); | |
318 | ASSERT(full_page_sectors_uptodate(src, i)); | |
4d100466 | 319 | steal_rbio_page(src, dest, i); |
4ae10b3a | 320 | } |
eb357060 QW |
321 | index_stripe_sectors(dest); |
322 | index_stripe_sectors(src); | |
4ae10b3a CM |
323 | } |
324 | ||
53b381b3 DW |
325 | /* |
326 | * merging means we take the bio_list from the victim and | |
327 | * splice it into the destination. The victim should | |
328 | * be discarded afterwards. | |
329 | * | |
330 | * must be called with dest->rbio_list_lock held | |
331 | */ | |
332 | static void merge_rbio(struct btrfs_raid_bio *dest, | |
333 | struct btrfs_raid_bio *victim) | |
334 | { | |
335 | bio_list_merge(&dest->bio_list, &victim->bio_list); | |
336 | dest->bio_list_bytes += victim->bio_list_bytes; | |
bd8f7e62 QW |
337 | /* Also inherit the bitmaps from @victim. */ |
338 | bitmap_or(&dest->dbitmap, &victim->dbitmap, &dest->dbitmap, | |
339 | dest->stripe_nsectors); | |
53b381b3 DW |
340 | bio_list_init(&victim->bio_list); |
341 | } | |
342 | ||
343 | /* | |
4ae10b3a CM |
344 | * used to prune items that are in the cache. The caller |
345 | * must hold the hash table lock. | |
346 | */ | |
347 | static void __remove_rbio_from_cache(struct btrfs_raid_bio *rbio) | |
348 | { | |
349 | int bucket = rbio_bucket(rbio); | |
350 | struct btrfs_stripe_hash_table *table; | |
351 | struct btrfs_stripe_hash *h; | |
352 | int freeit = 0; | |
353 | ||
354 | /* | |
355 | * check the bit again under the hash table lock. | |
356 | */ | |
357 | if (!test_bit(RBIO_CACHE_BIT, &rbio->flags)) | |
358 | return; | |
359 | ||
6a258d72 | 360 | table = rbio->bioc->fs_info->stripe_hash_table; |
4ae10b3a CM |
361 | h = table->table + bucket; |
362 | ||
363 | /* hold the lock for the bucket because we may be | |
364 | * removing it from the hash table | |
365 | */ | |
366 | spin_lock(&h->lock); | |
367 | ||
368 | /* | |
369 | * hold the lock for the bio list because we need | |
370 | * to make sure the bio list is empty | |
371 | */ | |
372 | spin_lock(&rbio->bio_list_lock); | |
373 | ||
374 | if (test_and_clear_bit(RBIO_CACHE_BIT, &rbio->flags)) { | |
375 | list_del_init(&rbio->stripe_cache); | |
376 | table->cache_size -= 1; | |
377 | freeit = 1; | |
378 | ||
379 | /* if the bio list isn't empty, this rbio is | |
380 | * still involved in an IO. We take it out | |
381 | * of the cache list, and drop the ref that | |
382 | * was held for the list. | |
383 | * | |
384 | * If the bio_list was empty, we also remove | |
385 | * the rbio from the hash_table, and drop | |
386 | * the corresponding ref | |
387 | */ | |
388 | if (bio_list_empty(&rbio->bio_list)) { | |
389 | if (!list_empty(&rbio->hash_list)) { | |
390 | list_del_init(&rbio->hash_list); | |
dec95574 | 391 | refcount_dec(&rbio->refs); |
4ae10b3a CM |
392 | BUG_ON(!list_empty(&rbio->plug_list)); |
393 | } | |
394 | } | |
395 | } | |
396 | ||
397 | spin_unlock(&rbio->bio_list_lock); | |
398 | spin_unlock(&h->lock); | |
399 | ||
400 | if (freeit) | |
ff2b64a2 | 401 | free_raid_bio(rbio); |
4ae10b3a CM |
402 | } |
403 | ||
404 | /* | |
405 | * prune a given rbio from the cache | |
406 | */ | |
407 | static void remove_rbio_from_cache(struct btrfs_raid_bio *rbio) | |
408 | { | |
409 | struct btrfs_stripe_hash_table *table; | |
410 | unsigned long flags; | |
411 | ||
412 | if (!test_bit(RBIO_CACHE_BIT, &rbio->flags)) | |
413 | return; | |
414 | ||
6a258d72 | 415 | table = rbio->bioc->fs_info->stripe_hash_table; |
4ae10b3a CM |
416 | |
417 | spin_lock_irqsave(&table->cache_lock, flags); | |
418 | __remove_rbio_from_cache(rbio); | |
419 | spin_unlock_irqrestore(&table->cache_lock, flags); | |
420 | } | |
421 | ||
422 | /* | |
423 | * remove everything in the cache | |
424 | */ | |
48a3b636 | 425 | static void btrfs_clear_rbio_cache(struct btrfs_fs_info *info) |
4ae10b3a CM |
426 | { |
427 | struct btrfs_stripe_hash_table *table; | |
428 | unsigned long flags; | |
429 | struct btrfs_raid_bio *rbio; | |
430 | ||
431 | table = info->stripe_hash_table; | |
432 | ||
433 | spin_lock_irqsave(&table->cache_lock, flags); | |
434 | while (!list_empty(&table->stripe_cache)) { | |
435 | rbio = list_entry(table->stripe_cache.next, | |
436 | struct btrfs_raid_bio, | |
437 | stripe_cache); | |
438 | __remove_rbio_from_cache(rbio); | |
439 | } | |
440 | spin_unlock_irqrestore(&table->cache_lock, flags); | |
441 | } | |
442 | ||
443 | /* | |
444 | * remove all cached entries and free the hash table | |
445 | * used by unmount | |
53b381b3 DW |
446 | */ |
447 | void btrfs_free_stripe_hash_table(struct btrfs_fs_info *info) | |
448 | { | |
449 | if (!info->stripe_hash_table) | |
450 | return; | |
4ae10b3a | 451 | btrfs_clear_rbio_cache(info); |
f749303b | 452 | kvfree(info->stripe_hash_table); |
53b381b3 DW |
453 | info->stripe_hash_table = NULL; |
454 | } | |
455 | ||
4ae10b3a CM |
456 | /* |
457 | * insert an rbio into the stripe cache. It | |
458 | * must have already been prepared by calling | |
459 | * cache_rbio_pages | |
460 | * | |
461 | * If this rbio was already cached, it gets | |
462 | * moved to the front of the lru. | |
463 | * | |
464 | * If the size of the rbio cache is too big, we | |
465 | * prune an item. | |
466 | */ | |
467 | static void cache_rbio(struct btrfs_raid_bio *rbio) | |
468 | { | |
469 | struct btrfs_stripe_hash_table *table; | |
470 | unsigned long flags; | |
471 | ||
472 | if (!test_bit(RBIO_CACHE_READY_BIT, &rbio->flags)) | |
473 | return; | |
474 | ||
6a258d72 | 475 | table = rbio->bioc->fs_info->stripe_hash_table; |
4ae10b3a CM |
476 | |
477 | spin_lock_irqsave(&table->cache_lock, flags); | |
478 | spin_lock(&rbio->bio_list_lock); | |
479 | ||
480 | /* bump our ref if we were not in the list before */ | |
481 | if (!test_and_set_bit(RBIO_CACHE_BIT, &rbio->flags)) | |
dec95574 | 482 | refcount_inc(&rbio->refs); |
4ae10b3a CM |
483 | |
484 | if (!list_empty(&rbio->stripe_cache)){ | |
485 | list_move(&rbio->stripe_cache, &table->stripe_cache); | |
486 | } else { | |
487 | list_add(&rbio->stripe_cache, &table->stripe_cache); | |
488 | table->cache_size += 1; | |
489 | } | |
490 | ||
491 | spin_unlock(&rbio->bio_list_lock); | |
492 | ||
493 | if (table->cache_size > RBIO_CACHE_SIZE) { | |
494 | struct btrfs_raid_bio *found; | |
495 | ||
496 | found = list_entry(table->stripe_cache.prev, | |
497 | struct btrfs_raid_bio, | |
498 | stripe_cache); | |
499 | ||
500 | if (found != rbio) | |
501 | __remove_rbio_from_cache(found); | |
502 | } | |
503 | ||
504 | spin_unlock_irqrestore(&table->cache_lock, flags); | |
4ae10b3a CM |
505 | } |
506 | ||
53b381b3 DW |
507 | /* |
508 | * helper function to run the xor_blocks api. It is only | |
509 | * able to do MAX_XOR_BLOCKS at a time, so we need to | |
510 | * loop through. | |
511 | */ | |
512 | static void run_xor(void **pages, int src_cnt, ssize_t len) | |
513 | { | |
514 | int src_off = 0; | |
515 | int xor_src_cnt = 0; | |
516 | void *dest = pages[src_cnt]; | |
517 | ||
518 | while(src_cnt > 0) { | |
519 | xor_src_cnt = min(src_cnt, MAX_XOR_BLOCKS); | |
520 | xor_blocks(xor_src_cnt, len, dest, pages + src_off); | |
521 | ||
522 | src_cnt -= xor_src_cnt; | |
523 | src_off += xor_src_cnt; | |
524 | } | |
525 | } | |
526 | ||
527 | /* | |
176571a1 DS |
528 | * Returns true if the bio list inside this rbio covers an entire stripe (no |
529 | * rmw required). | |
53b381b3 | 530 | */ |
176571a1 | 531 | static int rbio_is_full(struct btrfs_raid_bio *rbio) |
53b381b3 | 532 | { |
176571a1 | 533 | unsigned long flags; |
53b381b3 DW |
534 | unsigned long size = rbio->bio_list_bytes; |
535 | int ret = 1; | |
536 | ||
176571a1 | 537 | spin_lock_irqsave(&rbio->bio_list_lock, flags); |
ff18a4af | 538 | if (size != rbio->nr_data * BTRFS_STRIPE_LEN) |
53b381b3 | 539 | ret = 0; |
ff18a4af | 540 | BUG_ON(size > rbio->nr_data * BTRFS_STRIPE_LEN); |
53b381b3 | 541 | spin_unlock_irqrestore(&rbio->bio_list_lock, flags); |
176571a1 | 542 | |
53b381b3 DW |
543 | return ret; |
544 | } | |
545 | ||
546 | /* | |
547 | * returns 1 if it is safe to merge two rbios together. | |
548 | * The merging is safe if the two rbios correspond to | |
549 | * the same stripe and if they are both going in the same | |
550 | * direction (read vs write), and if neither one is | |
551 | * locked for final IO | |
552 | * | |
553 | * The caller is responsible for locking such that | |
554 | * rmw_locked is safe to test | |
555 | */ | |
556 | static int rbio_can_merge(struct btrfs_raid_bio *last, | |
557 | struct btrfs_raid_bio *cur) | |
558 | { | |
559 | if (test_bit(RBIO_RMW_LOCKED_BIT, &last->flags) || | |
560 | test_bit(RBIO_RMW_LOCKED_BIT, &cur->flags)) | |
561 | return 0; | |
562 | ||
4ae10b3a CM |
563 | /* |
564 | * we can't merge with cached rbios, since the | |
565 | * idea is that when we merge the destination | |
566 | * rbio is going to run our IO for us. We can | |
01327610 | 567 | * steal from cached rbios though, other functions |
4ae10b3a CM |
568 | * handle that. |
569 | */ | |
570 | if (test_bit(RBIO_CACHE_BIT, &last->flags) || | |
571 | test_bit(RBIO_CACHE_BIT, &cur->flags)) | |
572 | return 0; | |
573 | ||
4c664611 | 574 | if (last->bioc->raid_map[0] != cur->bioc->raid_map[0]) |
53b381b3 DW |
575 | return 0; |
576 | ||
5a6ac9ea MX |
577 | /* we can't merge with different operations */ |
578 | if (last->operation != cur->operation) | |
579 | return 0; | |
580 | /* | |
581 | * We've need read the full stripe from the drive. | |
582 | * check and repair the parity and write the new results. | |
583 | * | |
584 | * We're not allowed to add any new bios to the | |
585 | * bio list here, anyone else that wants to | |
586 | * change this stripe needs to do their own rmw. | |
587 | */ | |
db34be19 | 588 | if (last->operation == BTRFS_RBIO_PARITY_SCRUB) |
53b381b3 | 589 | return 0; |
53b381b3 | 590 | |
ad3daf1c QW |
591 | if (last->operation == BTRFS_RBIO_REBUILD_MISSING || |
592 | last->operation == BTRFS_RBIO_READ_REBUILD) | |
b4ee1782 OS |
593 | return 0; |
594 | ||
53b381b3 DW |
595 | return 1; |
596 | } | |
597 | ||
3e77605d QW |
598 | static unsigned int rbio_stripe_sector_index(const struct btrfs_raid_bio *rbio, |
599 | unsigned int stripe_nr, | |
600 | unsigned int sector_nr) | |
601 | { | |
602 | ASSERT(stripe_nr < rbio->real_stripes); | |
603 | ASSERT(sector_nr < rbio->stripe_nsectors); | |
604 | ||
605 | return stripe_nr * rbio->stripe_nsectors + sector_nr; | |
606 | } | |
607 | ||
608 | /* Return a sector from rbio->stripe_sectors, not from the bio list */ | |
609 | static struct sector_ptr *rbio_stripe_sector(const struct btrfs_raid_bio *rbio, | |
610 | unsigned int stripe_nr, | |
611 | unsigned int sector_nr) | |
612 | { | |
613 | return &rbio->stripe_sectors[rbio_stripe_sector_index(rbio, stripe_nr, | |
614 | sector_nr)]; | |
615 | } | |
616 | ||
1145059a QW |
617 | /* Grab a sector inside P stripe */ |
618 | static struct sector_ptr *rbio_pstripe_sector(const struct btrfs_raid_bio *rbio, | |
619 | unsigned int sector_nr) | |
b7178a5f | 620 | { |
1145059a | 621 | return rbio_stripe_sector(rbio, rbio->nr_data, sector_nr); |
b7178a5f ZL |
622 | } |
623 | ||
1145059a QW |
624 | /* Grab a sector inside Q stripe, return NULL if not RAID6 */ |
625 | static struct sector_ptr *rbio_qstripe_sector(const struct btrfs_raid_bio *rbio, | |
626 | unsigned int sector_nr) | |
53b381b3 | 627 | { |
1145059a QW |
628 | if (rbio->nr_data + 1 == rbio->real_stripes) |
629 | return NULL; | |
630 | return rbio_stripe_sector(rbio, rbio->nr_data + 1, sector_nr); | |
53b381b3 DW |
631 | } |
632 | ||
53b381b3 DW |
633 | /* |
634 | * The first stripe in the table for a logical address | |
635 | * has the lock. rbios are added in one of three ways: | |
636 | * | |
637 | * 1) Nobody has the stripe locked yet. The rbio is given | |
638 | * the lock and 0 is returned. The caller must start the IO | |
639 | * themselves. | |
640 | * | |
641 | * 2) Someone has the stripe locked, but we're able to merge | |
642 | * with the lock owner. The rbio is freed and the IO will | |
643 | * start automatically along with the existing rbio. 1 is returned. | |
644 | * | |
645 | * 3) Someone has the stripe locked, but we're not able to merge. | |
646 | * The rbio is added to the lock owner's plug list, or merged into | |
647 | * an rbio already on the plug list. When the lock owner unlocks, | |
648 | * the next rbio on the list is run and the IO is started automatically. | |
649 | * 1 is returned | |
650 | * | |
651 | * If we return 0, the caller still owns the rbio and must continue with | |
652 | * IO submission. If we return 1, the caller must assume the rbio has | |
653 | * already been freed. | |
654 | */ | |
655 | static noinline int lock_stripe_add(struct btrfs_raid_bio *rbio) | |
656 | { | |
721860d5 | 657 | struct btrfs_stripe_hash *h; |
53b381b3 DW |
658 | struct btrfs_raid_bio *cur; |
659 | struct btrfs_raid_bio *pending; | |
660 | unsigned long flags; | |
53b381b3 | 661 | struct btrfs_raid_bio *freeit = NULL; |
4ae10b3a | 662 | struct btrfs_raid_bio *cache_drop = NULL; |
53b381b3 | 663 | int ret = 0; |
53b381b3 | 664 | |
6a258d72 | 665 | h = rbio->bioc->fs_info->stripe_hash_table->table + rbio_bucket(rbio); |
721860d5 | 666 | |
53b381b3 DW |
667 | spin_lock_irqsave(&h->lock, flags); |
668 | list_for_each_entry(cur, &h->hash_list, hash_list) { | |
4c664611 | 669 | if (cur->bioc->raid_map[0] != rbio->bioc->raid_map[0]) |
9d6cb1b0 | 670 | continue; |
4ae10b3a | 671 | |
9d6cb1b0 | 672 | spin_lock(&cur->bio_list_lock); |
4ae10b3a | 673 | |
9d6cb1b0 JT |
674 | /* Can we steal this cached rbio's pages? */ |
675 | if (bio_list_empty(&cur->bio_list) && | |
676 | list_empty(&cur->plug_list) && | |
677 | test_bit(RBIO_CACHE_BIT, &cur->flags) && | |
678 | !test_bit(RBIO_RMW_LOCKED_BIT, &cur->flags)) { | |
679 | list_del_init(&cur->hash_list); | |
680 | refcount_dec(&cur->refs); | |
53b381b3 | 681 | |
9d6cb1b0 JT |
682 | steal_rbio(cur, rbio); |
683 | cache_drop = cur; | |
684 | spin_unlock(&cur->bio_list_lock); | |
4ae10b3a | 685 | |
9d6cb1b0 JT |
686 | goto lockit; |
687 | } | |
53b381b3 | 688 | |
9d6cb1b0 JT |
689 | /* Can we merge into the lock owner? */ |
690 | if (rbio_can_merge(cur, rbio)) { | |
691 | merge_rbio(cur, rbio); | |
53b381b3 | 692 | spin_unlock(&cur->bio_list_lock); |
9d6cb1b0 | 693 | freeit = rbio; |
53b381b3 DW |
694 | ret = 1; |
695 | goto out; | |
696 | } | |
9d6cb1b0 JT |
697 | |
698 | ||
699 | /* | |
700 | * We couldn't merge with the running rbio, see if we can merge | |
701 | * with the pending ones. We don't have to check for rmw_locked | |
702 | * because there is no way they are inside finish_rmw right now | |
703 | */ | |
704 | list_for_each_entry(pending, &cur->plug_list, plug_list) { | |
705 | if (rbio_can_merge(pending, rbio)) { | |
706 | merge_rbio(pending, rbio); | |
707 | spin_unlock(&cur->bio_list_lock); | |
708 | freeit = rbio; | |
709 | ret = 1; | |
710 | goto out; | |
711 | } | |
712 | } | |
713 | ||
714 | /* | |
715 | * No merging, put us on the tail of the plug list, our rbio | |
716 | * will be started with the currently running rbio unlocks | |
717 | */ | |
718 | list_add_tail(&rbio->plug_list, &cur->plug_list); | |
719 | spin_unlock(&cur->bio_list_lock); | |
720 | ret = 1; | |
721 | goto out; | |
53b381b3 | 722 | } |
4ae10b3a | 723 | lockit: |
dec95574 | 724 | refcount_inc(&rbio->refs); |
53b381b3 DW |
725 | list_add(&rbio->hash_list, &h->hash_list); |
726 | out: | |
727 | spin_unlock_irqrestore(&h->lock, flags); | |
4ae10b3a CM |
728 | if (cache_drop) |
729 | remove_rbio_from_cache(cache_drop); | |
53b381b3 | 730 | if (freeit) |
ff2b64a2 | 731 | free_raid_bio(freeit); |
53b381b3 DW |
732 | return ret; |
733 | } | |
734 | ||
d817ce35 QW |
735 | static void recover_rbio_work_locked(struct work_struct *work); |
736 | ||
53b381b3 DW |
737 | /* |
738 | * called as rmw or parity rebuild is completed. If the plug list has more | |
739 | * rbios waiting for this stripe, the next one on the list will be started | |
740 | */ | |
741 | static noinline void unlock_stripe(struct btrfs_raid_bio *rbio) | |
742 | { | |
743 | int bucket; | |
744 | struct btrfs_stripe_hash *h; | |
745 | unsigned long flags; | |
4ae10b3a | 746 | int keep_cache = 0; |
53b381b3 DW |
747 | |
748 | bucket = rbio_bucket(rbio); | |
6a258d72 | 749 | h = rbio->bioc->fs_info->stripe_hash_table->table + bucket; |
53b381b3 | 750 | |
4ae10b3a CM |
751 | if (list_empty(&rbio->plug_list)) |
752 | cache_rbio(rbio); | |
753 | ||
53b381b3 DW |
754 | spin_lock_irqsave(&h->lock, flags); |
755 | spin_lock(&rbio->bio_list_lock); | |
756 | ||
757 | if (!list_empty(&rbio->hash_list)) { | |
4ae10b3a CM |
758 | /* |
759 | * if we're still cached and there is no other IO | |
760 | * to perform, just leave this rbio here for others | |
761 | * to steal from later | |
762 | */ | |
763 | if (list_empty(&rbio->plug_list) && | |
764 | test_bit(RBIO_CACHE_BIT, &rbio->flags)) { | |
765 | keep_cache = 1; | |
766 | clear_bit(RBIO_RMW_LOCKED_BIT, &rbio->flags); | |
767 | BUG_ON(!bio_list_empty(&rbio->bio_list)); | |
768 | goto done; | |
769 | } | |
53b381b3 DW |
770 | |
771 | list_del_init(&rbio->hash_list); | |
dec95574 | 772 | refcount_dec(&rbio->refs); |
53b381b3 DW |
773 | |
774 | /* | |
775 | * we use the plug list to hold all the rbios | |
776 | * waiting for the chance to lock this stripe. | |
777 | * hand the lock over to one of them. | |
778 | */ | |
779 | if (!list_empty(&rbio->plug_list)) { | |
780 | struct btrfs_raid_bio *next; | |
781 | struct list_head *head = rbio->plug_list.next; | |
782 | ||
783 | next = list_entry(head, struct btrfs_raid_bio, | |
784 | plug_list); | |
785 | ||
786 | list_del_init(&rbio->plug_list); | |
787 | ||
788 | list_add(&next->hash_list, &h->hash_list); | |
dec95574 | 789 | refcount_inc(&next->refs); |
53b381b3 DW |
790 | spin_unlock(&rbio->bio_list_lock); |
791 | spin_unlock_irqrestore(&h->lock, flags); | |
792 | ||
1b94b556 | 793 | if (next->operation == BTRFS_RBIO_READ_REBUILD) |
d817ce35 | 794 | start_async_work(next, recover_rbio_work_locked); |
b4ee1782 OS |
795 | else if (next->operation == BTRFS_RBIO_REBUILD_MISSING) { |
796 | steal_rbio(rbio, next); | |
d817ce35 | 797 | start_async_work(next, recover_rbio_work_locked); |
b4ee1782 | 798 | } else if (next->operation == BTRFS_RBIO_WRITE) { |
4ae10b3a | 799 | steal_rbio(rbio, next); |
93723095 | 800 | start_async_work(next, rmw_rbio_work_locked); |
5a6ac9ea MX |
801 | } else if (next->operation == BTRFS_RBIO_PARITY_SCRUB) { |
802 | steal_rbio(rbio, next); | |
6bfd0133 | 803 | start_async_work(next, scrub_rbio_work_locked); |
4ae10b3a | 804 | } |
53b381b3 DW |
805 | |
806 | goto done_nolock; | |
53b381b3 DW |
807 | } |
808 | } | |
4ae10b3a | 809 | done: |
53b381b3 DW |
810 | spin_unlock(&rbio->bio_list_lock); |
811 | spin_unlock_irqrestore(&h->lock, flags); | |
812 | ||
813 | done_nolock: | |
4ae10b3a CM |
814 | if (!keep_cache) |
815 | remove_rbio_from_cache(rbio); | |
53b381b3 DW |
816 | } |
817 | ||
7583d8d0 | 818 | static void rbio_endio_bio_list(struct bio *cur, blk_status_t err) |
53b381b3 | 819 | { |
7583d8d0 LB |
820 | struct bio *next; |
821 | ||
822 | while (cur) { | |
823 | next = cur->bi_next; | |
824 | cur->bi_next = NULL; | |
825 | cur->bi_status = err; | |
826 | bio_endio(cur); | |
827 | cur = next; | |
828 | } | |
53b381b3 DW |
829 | } |
830 | ||
831 | /* | |
832 | * this frees the rbio and runs through all the bios in the | |
833 | * bio_list and calls end_io on them | |
834 | */ | |
4e4cbee9 | 835 | static void rbio_orig_end_io(struct btrfs_raid_bio *rbio, blk_status_t err) |
53b381b3 DW |
836 | { |
837 | struct bio *cur = bio_list_get(&rbio->bio_list); | |
7583d8d0 | 838 | struct bio *extra; |
4245215d | 839 | |
c5a41562 QW |
840 | kfree(rbio->csum_buf); |
841 | bitmap_free(rbio->csum_bitmap); | |
842 | rbio->csum_buf = NULL; | |
843 | rbio->csum_bitmap = NULL; | |
844 | ||
bd8f7e62 QW |
845 | /* |
846 | * Clear the data bitmap, as the rbio may be cached for later usage. | |
847 | * do this before before unlock_stripe() so there will be no new bio | |
848 | * for this bio. | |
849 | */ | |
850 | bitmap_clear(&rbio->dbitmap, 0, rbio->stripe_nsectors); | |
4245215d | 851 | |
7583d8d0 LB |
852 | /* |
853 | * At this moment, rbio->bio_list is empty, however since rbio does not | |
854 | * always have RBIO_RMW_LOCKED_BIT set and rbio is still linked on the | |
855 | * hash list, rbio may be merged with others so that rbio->bio_list | |
856 | * becomes non-empty. | |
857 | * Once unlock_stripe() is done, rbio->bio_list will not be updated any | |
858 | * more and we can call bio_endio() on all queued bios. | |
859 | */ | |
860 | unlock_stripe(rbio); | |
861 | extra = bio_list_get(&rbio->bio_list); | |
ff2b64a2 | 862 | free_raid_bio(rbio); |
53b381b3 | 863 | |
7583d8d0 LB |
864 | rbio_endio_bio_list(cur, err); |
865 | if (extra) | |
866 | rbio_endio_bio_list(extra, err); | |
53b381b3 DW |
867 | } |
868 | ||
43dd529a DS |
869 | /* |
870 | * Get a sector pointer specified by its @stripe_nr and @sector_nr. | |
3e77605d QW |
871 | * |
872 | * @rbio: The raid bio | |
873 | * @stripe_nr: Stripe number, valid range [0, real_stripe) | |
874 | * @sector_nr: Sector number inside the stripe, | |
875 | * valid range [0, stripe_nsectors) | |
876 | * @bio_list_only: Whether to use sectors inside the bio list only. | |
877 | * | |
878 | * The read/modify/write code wants to reuse the original bio page as much | |
879 | * as possible, and only use stripe_sectors as fallback. | |
880 | */ | |
881 | static struct sector_ptr *sector_in_rbio(struct btrfs_raid_bio *rbio, | |
882 | int stripe_nr, int sector_nr, | |
883 | bool bio_list_only) | |
884 | { | |
885 | struct sector_ptr *sector; | |
886 | int index; | |
887 | ||
888 | ASSERT(stripe_nr >= 0 && stripe_nr < rbio->real_stripes); | |
889 | ASSERT(sector_nr >= 0 && sector_nr < rbio->stripe_nsectors); | |
890 | ||
891 | index = stripe_nr * rbio->stripe_nsectors + sector_nr; | |
892 | ASSERT(index >= 0 && index < rbio->nr_sectors); | |
893 | ||
894 | spin_lock_irq(&rbio->bio_list_lock); | |
895 | sector = &rbio->bio_sectors[index]; | |
896 | if (sector->page || bio_list_only) { | |
897 | /* Don't return sector without a valid page pointer */ | |
898 | if (!sector->page) | |
899 | sector = NULL; | |
900 | spin_unlock_irq(&rbio->bio_list_lock); | |
901 | return sector; | |
902 | } | |
903 | spin_unlock_irq(&rbio->bio_list_lock); | |
904 | ||
905 | return &rbio->stripe_sectors[index]; | |
906 | } | |
907 | ||
53b381b3 DW |
908 | /* |
909 | * allocation and initial setup for the btrfs_raid_bio. Not | |
910 | * this does not allocate any pages for rbio->pages. | |
911 | */ | |
2ff7e61e | 912 | static struct btrfs_raid_bio *alloc_rbio(struct btrfs_fs_info *fs_info, |
ff18a4af | 913 | struct btrfs_io_context *bioc) |
53b381b3 | 914 | { |
843de58b | 915 | const unsigned int real_stripes = bioc->num_stripes - bioc->num_tgtdevs; |
ff18a4af | 916 | const unsigned int stripe_npages = BTRFS_STRIPE_LEN >> PAGE_SHIFT; |
843de58b | 917 | const unsigned int num_pages = stripe_npages * real_stripes; |
ff18a4af CH |
918 | const unsigned int stripe_nsectors = |
919 | BTRFS_STRIPE_LEN >> fs_info->sectorsize_bits; | |
94efbe19 | 920 | const unsigned int num_sectors = stripe_nsectors * real_stripes; |
53b381b3 | 921 | struct btrfs_raid_bio *rbio; |
53b381b3 | 922 | |
94efbe19 QW |
923 | /* PAGE_SIZE must also be aligned to sectorsize for subpage support */ |
924 | ASSERT(IS_ALIGNED(PAGE_SIZE, fs_info->sectorsize)); | |
c67c68eb QW |
925 | /* |
926 | * Our current stripe len should be fixed to 64k thus stripe_nsectors | |
927 | * (at most 16) should be no larger than BITS_PER_LONG. | |
928 | */ | |
929 | ASSERT(stripe_nsectors <= BITS_PER_LONG); | |
843de58b | 930 | |
797d74b7 | 931 | rbio = kzalloc(sizeof(*rbio), GFP_NOFS); |
af8e2d1d | 932 | if (!rbio) |
53b381b3 | 933 | return ERR_PTR(-ENOMEM); |
797d74b7 QW |
934 | rbio->stripe_pages = kcalloc(num_pages, sizeof(struct page *), |
935 | GFP_NOFS); | |
936 | rbio->bio_sectors = kcalloc(num_sectors, sizeof(struct sector_ptr), | |
937 | GFP_NOFS); | |
938 | rbio->stripe_sectors = kcalloc(num_sectors, sizeof(struct sector_ptr), | |
939 | GFP_NOFS); | |
940 | rbio->finish_pointers = kcalloc(real_stripes, sizeof(void *), GFP_NOFS); | |
2942a50d | 941 | rbio->error_bitmap = bitmap_zalloc(num_sectors, GFP_NOFS); |
797d74b7 QW |
942 | |
943 | if (!rbio->stripe_pages || !rbio->bio_sectors || !rbio->stripe_sectors || | |
2942a50d | 944 | !rbio->finish_pointers || !rbio->error_bitmap) { |
797d74b7 QW |
945 | free_raid_bio_pointers(rbio); |
946 | kfree(rbio); | |
947 | return ERR_PTR(-ENOMEM); | |
948 | } | |
53b381b3 DW |
949 | |
950 | bio_list_init(&rbio->bio_list); | |
d817ce35 | 951 | init_waitqueue_head(&rbio->io_wait); |
53b381b3 DW |
952 | INIT_LIST_HEAD(&rbio->plug_list); |
953 | spin_lock_init(&rbio->bio_list_lock); | |
4ae10b3a | 954 | INIT_LIST_HEAD(&rbio->stripe_cache); |
53b381b3 | 955 | INIT_LIST_HEAD(&rbio->hash_list); |
f1c29379 | 956 | btrfs_get_bioc(bioc); |
4c664611 | 957 | rbio->bioc = bioc; |
53b381b3 | 958 | rbio->nr_pages = num_pages; |
94efbe19 | 959 | rbio->nr_sectors = num_sectors; |
2c8cdd6e | 960 | rbio->real_stripes = real_stripes; |
5a6ac9ea | 961 | rbio->stripe_npages = stripe_npages; |
94efbe19 | 962 | rbio->stripe_nsectors = stripe_nsectors; |
dec95574 | 963 | refcount_set(&rbio->refs, 1); |
b89e1b01 | 964 | atomic_set(&rbio->stripes_pending, 0); |
53b381b3 | 965 | |
0b30f719 QW |
966 | ASSERT(btrfs_nr_parity_stripes(bioc->map_type)); |
967 | rbio->nr_data = real_stripes - btrfs_nr_parity_stripes(bioc->map_type); | |
53b381b3 | 968 | |
53b381b3 DW |
969 | return rbio; |
970 | } | |
971 | ||
972 | /* allocate pages for all the stripes in the bio, including parity */ | |
973 | static int alloc_rbio_pages(struct btrfs_raid_bio *rbio) | |
974 | { | |
eb357060 QW |
975 | int ret; |
976 | ||
977 | ret = btrfs_alloc_page_array(rbio->nr_pages, rbio->stripe_pages); | |
978 | if (ret < 0) | |
979 | return ret; | |
980 | /* Mapping all sectors */ | |
981 | index_stripe_sectors(rbio); | |
982 | return 0; | |
53b381b3 DW |
983 | } |
984 | ||
b7178a5f | 985 | /* only allocate pages for p/q stripes */ |
53b381b3 DW |
986 | static int alloc_rbio_parity_pages(struct btrfs_raid_bio *rbio) |
987 | { | |
f77183dc | 988 | const int data_pages = rbio->nr_data * rbio->stripe_npages; |
eb357060 | 989 | int ret; |
53b381b3 | 990 | |
eb357060 QW |
991 | ret = btrfs_alloc_page_array(rbio->nr_pages - data_pages, |
992 | rbio->stripe_pages + data_pages); | |
993 | if (ret < 0) | |
994 | return ret; | |
995 | ||
996 | index_stripe_sectors(rbio); | |
997 | return 0; | |
53b381b3 DW |
998 | } |
999 | ||
75b47033 | 1000 | /* |
67da05b3 | 1001 | * Return the total number of errors found in the vertical stripe of @sector_nr. |
75b47033 QW |
1002 | * |
1003 | * @faila and @failb will also be updated to the first and second stripe | |
1004 | * number of the errors. | |
1005 | */ | |
1006 | static int get_rbio_veritical_errors(struct btrfs_raid_bio *rbio, int sector_nr, | |
1007 | int *faila, int *failb) | |
1008 | { | |
1009 | int stripe_nr; | |
1010 | int found_errors = 0; | |
1011 | ||
ad3daf1c QW |
1012 | if (faila || failb) { |
1013 | /* | |
1014 | * Both @faila and @failb should be valid pointers if any of | |
1015 | * them is specified. | |
1016 | */ | |
1017 | ASSERT(faila && failb); | |
1018 | *faila = -1; | |
1019 | *failb = -1; | |
1020 | } | |
75b47033 QW |
1021 | |
1022 | for (stripe_nr = 0; stripe_nr < rbio->real_stripes; stripe_nr++) { | |
1023 | int total_sector_nr = stripe_nr * rbio->stripe_nsectors + sector_nr; | |
1024 | ||
1025 | if (test_bit(total_sector_nr, rbio->error_bitmap)) { | |
1026 | found_errors++; | |
ad3daf1c QW |
1027 | if (faila) { |
1028 | /* Update faila and failb. */ | |
1029 | if (*faila < 0) | |
1030 | *faila = stripe_nr; | |
1031 | else if (*failb < 0) | |
1032 | *failb = stripe_nr; | |
1033 | } | |
75b47033 QW |
1034 | } |
1035 | } | |
1036 | return found_errors; | |
1037 | } | |
1038 | ||
53b381b3 | 1039 | /* |
3e77605d QW |
1040 | * Add a single sector @sector into our list of bios for IO. |
1041 | * | |
1042 | * Return 0 if everything went well. | |
1043 | * Return <0 for error. | |
53b381b3 | 1044 | */ |
3e77605d QW |
1045 | static int rbio_add_io_sector(struct btrfs_raid_bio *rbio, |
1046 | struct bio_list *bio_list, | |
1047 | struct sector_ptr *sector, | |
1048 | unsigned int stripe_nr, | |
1049 | unsigned int sector_nr, | |
bf9486d6 | 1050 | enum req_op op) |
53b381b3 | 1051 | { |
3e77605d | 1052 | const u32 sectorsize = rbio->bioc->fs_info->sectorsize; |
53b381b3 | 1053 | struct bio *last = bio_list->tail; |
53b381b3 DW |
1054 | int ret; |
1055 | struct bio *bio; | |
4c664611 | 1056 | struct btrfs_io_stripe *stripe; |
53b381b3 DW |
1057 | u64 disk_start; |
1058 | ||
3e77605d QW |
1059 | /* |
1060 | * Note: here stripe_nr has taken device replace into consideration, | |
1061 | * thus it can be larger than rbio->real_stripe. | |
1062 | * So here we check against bioc->num_stripes, not rbio->real_stripes. | |
1063 | */ | |
1064 | ASSERT(stripe_nr >= 0 && stripe_nr < rbio->bioc->num_stripes); | |
1065 | ASSERT(sector_nr >= 0 && sector_nr < rbio->stripe_nsectors); | |
1066 | ASSERT(sector->page); | |
1067 | ||
4c664611 | 1068 | stripe = &rbio->bioc->stripes[stripe_nr]; |
3e77605d | 1069 | disk_start = stripe->physical + sector_nr * sectorsize; |
53b381b3 DW |
1070 | |
1071 | /* if the device is missing, just fail this stripe */ | |
2942a50d | 1072 | if (!stripe->dev->bdev) { |
ad3daf1c QW |
1073 | int found_errors; |
1074 | ||
2942a50d QW |
1075 | set_bit(stripe_nr * rbio->stripe_nsectors + sector_nr, |
1076 | rbio->error_bitmap); | |
ad3daf1c QW |
1077 | |
1078 | /* Check if we have reached tolerance early. */ | |
1079 | found_errors = get_rbio_veritical_errors(rbio, sector_nr, | |
1080 | NULL, NULL); | |
1081 | if (found_errors > rbio->bioc->max_errors) | |
1082 | return -EIO; | |
1083 | return 0; | |
2942a50d | 1084 | } |
53b381b3 DW |
1085 | |
1086 | /* see if we can add this page onto our existing bio */ | |
1087 | if (last) { | |
1201b58b | 1088 | u64 last_end = last->bi_iter.bi_sector << 9; |
4f024f37 | 1089 | last_end += last->bi_iter.bi_size; |
53b381b3 DW |
1090 | |
1091 | /* | |
1092 | * we can't merge these if they are from different | |
1093 | * devices or if they are not contiguous | |
1094 | */ | |
f90ae76a | 1095 | if (last_end == disk_start && !last->bi_status && |
309dca30 | 1096 | last->bi_bdev == stripe->dev->bdev) { |
3e77605d QW |
1097 | ret = bio_add_page(last, sector->page, sectorsize, |
1098 | sector->pgoff); | |
1099 | if (ret == sectorsize) | |
53b381b3 DW |
1100 | return 0; |
1101 | } | |
1102 | } | |
1103 | ||
1104 | /* put a new bio on the list */ | |
ff18a4af CH |
1105 | bio = bio_alloc(stripe->dev->bdev, |
1106 | max(BTRFS_STRIPE_LEN >> PAGE_SHIFT, 1), | |
bf9486d6 | 1107 | op, GFP_NOFS); |
4f024f37 | 1108 | bio->bi_iter.bi_sector = disk_start >> 9; |
e01bf588 | 1109 | bio->bi_private = rbio; |
53b381b3 | 1110 | |
3e77605d | 1111 | bio_add_page(bio, sector->page, sectorsize, sector->pgoff); |
53b381b3 DW |
1112 | bio_list_add(bio_list, bio); |
1113 | return 0; | |
1114 | } | |
1115 | ||
00425dd9 QW |
1116 | static void index_one_bio(struct btrfs_raid_bio *rbio, struct bio *bio) |
1117 | { | |
1118 | const u32 sectorsize = rbio->bioc->fs_info->sectorsize; | |
1119 | struct bio_vec bvec; | |
1120 | struct bvec_iter iter; | |
1121 | u32 offset = (bio->bi_iter.bi_sector << SECTOR_SHIFT) - | |
1122 | rbio->bioc->raid_map[0]; | |
1123 | ||
00425dd9 QW |
1124 | bio_for_each_segment(bvec, bio, iter) { |
1125 | u32 bvec_offset; | |
1126 | ||
1127 | for (bvec_offset = 0; bvec_offset < bvec.bv_len; | |
1128 | bvec_offset += sectorsize, offset += sectorsize) { | |
1129 | int index = offset / sectorsize; | |
1130 | struct sector_ptr *sector = &rbio->bio_sectors[index]; | |
1131 | ||
1132 | sector->page = bvec.bv_page; | |
1133 | sector->pgoff = bvec.bv_offset + bvec_offset; | |
1134 | ASSERT(sector->pgoff < PAGE_SIZE); | |
1135 | } | |
1136 | } | |
1137 | } | |
1138 | ||
53b381b3 DW |
1139 | /* |
1140 | * helper function to walk our bio list and populate the bio_pages array with | |
1141 | * the result. This seems expensive, but it is faster than constantly | |
1142 | * searching through the bio list as we setup the IO in finish_rmw or stripe | |
1143 | * reconstruction. | |
1144 | * | |
1145 | * This must be called before you trust the answers from page_in_rbio | |
1146 | */ | |
1147 | static void index_rbio_pages(struct btrfs_raid_bio *rbio) | |
1148 | { | |
1149 | struct bio *bio; | |
53b381b3 DW |
1150 | |
1151 | spin_lock_irq(&rbio->bio_list_lock); | |
00425dd9 QW |
1152 | bio_list_for_each(bio, &rbio->bio_list) |
1153 | index_one_bio(rbio, bio); | |
1154 | ||
53b381b3 DW |
1155 | spin_unlock_irq(&rbio->bio_list_lock); |
1156 | } | |
1157 | ||
b8bea09a QW |
1158 | static void bio_get_trace_info(struct btrfs_raid_bio *rbio, struct bio *bio, |
1159 | struct raid56_bio_trace_info *trace_info) | |
1160 | { | |
1161 | const struct btrfs_io_context *bioc = rbio->bioc; | |
1162 | int i; | |
1163 | ||
1164 | ASSERT(bioc); | |
1165 | ||
1166 | /* We rely on bio->bi_bdev to find the stripe number. */ | |
1167 | if (!bio->bi_bdev) | |
1168 | goto not_found; | |
1169 | ||
1170 | for (i = 0; i < bioc->num_stripes; i++) { | |
1171 | if (bio->bi_bdev != bioc->stripes[i].dev->bdev) | |
1172 | continue; | |
1173 | trace_info->stripe_nr = i; | |
1174 | trace_info->devid = bioc->stripes[i].dev->devid; | |
1175 | trace_info->offset = (bio->bi_iter.bi_sector << SECTOR_SHIFT) - | |
1176 | bioc->stripes[i].physical; | |
1177 | return; | |
1178 | } | |
1179 | ||
1180 | not_found: | |
1181 | trace_info->devid = -1; | |
1182 | trace_info->offset = -1; | |
1183 | trace_info->stripe_nr = -1; | |
1184 | } | |
1185 | ||
801fcfc5 CH |
1186 | static inline void bio_list_put(struct bio_list *bio_list) |
1187 | { | |
1188 | struct bio *bio; | |
1189 | ||
1190 | while ((bio = bio_list_pop(bio_list))) | |
1191 | bio_put(bio); | |
1192 | } | |
1193 | ||
67da05b3 | 1194 | /* Generate PQ for one vertical stripe. */ |
30e3c897 QW |
1195 | static void generate_pq_vertical(struct btrfs_raid_bio *rbio, int sectornr) |
1196 | { | |
1197 | void **pointers = rbio->finish_pointers; | |
1198 | const u32 sectorsize = rbio->bioc->fs_info->sectorsize; | |
1199 | struct sector_ptr *sector; | |
1200 | int stripe; | |
1201 | const bool has_qstripe = rbio->bioc->map_type & BTRFS_BLOCK_GROUP_RAID6; | |
1202 | ||
1203 | /* First collect one sector from each data stripe */ | |
1204 | for (stripe = 0; stripe < rbio->nr_data; stripe++) { | |
1205 | sector = sector_in_rbio(rbio, stripe, sectornr, 0); | |
1206 | pointers[stripe] = kmap_local_page(sector->page) + | |
1207 | sector->pgoff; | |
1208 | } | |
1209 | ||
1210 | /* Then add the parity stripe */ | |
1211 | sector = rbio_pstripe_sector(rbio, sectornr); | |
1212 | sector->uptodate = 1; | |
1213 | pointers[stripe++] = kmap_local_page(sector->page) + sector->pgoff; | |
1214 | ||
1215 | if (has_qstripe) { | |
1216 | /* | |
1217 | * RAID6, add the qstripe and call the library function | |
1218 | * to fill in our p/q | |
1219 | */ | |
1220 | sector = rbio_qstripe_sector(rbio, sectornr); | |
1221 | sector->uptodate = 1; | |
1222 | pointers[stripe++] = kmap_local_page(sector->page) + | |
1223 | sector->pgoff; | |
1224 | ||
1225 | raid6_call.gen_syndrome(rbio->real_stripes, sectorsize, | |
1226 | pointers); | |
1227 | } else { | |
1228 | /* raid5 */ | |
1229 | memcpy(pointers[rbio->nr_data], pointers[0], sectorsize); | |
1230 | run_xor(pointers + 1, rbio->nr_data - 1, sectorsize); | |
1231 | } | |
1232 | for (stripe = stripe - 1; stripe >= 0; stripe--) | |
1233 | kunmap_local(pointers[stripe]); | |
1234 | } | |
1235 | ||
6486d21c QW |
1236 | static int rmw_assemble_write_bios(struct btrfs_raid_bio *rbio, |
1237 | struct bio_list *bio_list) | |
53b381b3 | 1238 | { |
36920044 QW |
1239 | /* The total sector number inside the full stripe. */ |
1240 | int total_sector_nr; | |
3e77605d | 1241 | int sectornr; |
6486d21c | 1242 | int stripe; |
53b381b3 DW |
1243 | int ret; |
1244 | ||
6486d21c | 1245 | ASSERT(bio_list_size(bio_list) == 0); |
53b381b3 | 1246 | |
bd8f7e62 QW |
1247 | /* We should have at least one data sector. */ |
1248 | ASSERT(bitmap_weight(&rbio->dbitmap, rbio->stripe_nsectors)); | |
1249 | ||
5eb30ee2 QW |
1250 | /* |
1251 | * Reset errors, as we may have errors inherited from from degraded | |
1252 | * write. | |
1253 | */ | |
2942a50d | 1254 | bitmap_clear(rbio->error_bitmap, 0, rbio->nr_sectors); |
5eb30ee2 | 1255 | |
53b381b3 | 1256 | /* |
6486d21c | 1257 | * Start assembly. Make bios for everything from the higher layers (the |
36920044 | 1258 | * bio_list in our rbio) and our P/Q. Ignore everything else. |
53b381b3 | 1259 | */ |
36920044 QW |
1260 | for (total_sector_nr = 0; total_sector_nr < rbio->nr_sectors; |
1261 | total_sector_nr++) { | |
1262 | struct sector_ptr *sector; | |
3e77605d | 1263 | |
36920044 QW |
1264 | stripe = total_sector_nr / rbio->stripe_nsectors; |
1265 | sectornr = total_sector_nr % rbio->stripe_nsectors; | |
53b381b3 | 1266 | |
36920044 QW |
1267 | /* This vertical stripe has no data, skip it. */ |
1268 | if (!test_bit(sectornr, &rbio->dbitmap)) | |
1269 | continue; | |
53b381b3 | 1270 | |
36920044 QW |
1271 | if (stripe < rbio->nr_data) { |
1272 | sector = sector_in_rbio(rbio, stripe, sectornr, 1); | |
1273 | if (!sector) | |
1274 | continue; | |
1275 | } else { | |
1276 | sector = rbio_stripe_sector(rbio, stripe, sectornr); | |
53b381b3 | 1277 | } |
36920044 | 1278 | |
6486d21c | 1279 | ret = rbio_add_io_sector(rbio, bio_list, sector, stripe, |
ff18a4af | 1280 | sectornr, REQ_OP_WRITE); |
36920044 | 1281 | if (ret) |
6486d21c | 1282 | goto error; |
53b381b3 DW |
1283 | } |
1284 | ||
6486d21c QW |
1285 | if (likely(!rbio->bioc->num_tgtdevs)) |
1286 | return 0; | |
2c8cdd6e | 1287 | |
6486d21c | 1288 | /* Make a copy for the replace target device. */ |
36920044 QW |
1289 | for (total_sector_nr = 0; total_sector_nr < rbio->nr_sectors; |
1290 | total_sector_nr++) { | |
1291 | struct sector_ptr *sector; | |
2c8cdd6e | 1292 | |
36920044 QW |
1293 | stripe = total_sector_nr / rbio->stripe_nsectors; |
1294 | sectornr = total_sector_nr % rbio->stripe_nsectors; | |
3e77605d | 1295 | |
6486d21c | 1296 | if (!rbio->bioc->tgtdev_map[stripe]) { |
36920044 QW |
1297 | /* |
1298 | * We can skip the whole stripe completely, note | |
1299 | * total_sector_nr will be increased by one anyway. | |
1300 | */ | |
1301 | ASSERT(sectornr == 0); | |
1302 | total_sector_nr += rbio->stripe_nsectors - 1; | |
1303 | continue; | |
1304 | } | |
2c8cdd6e | 1305 | |
36920044 QW |
1306 | /* This vertical stripe has no data, skip it. */ |
1307 | if (!test_bit(sectornr, &rbio->dbitmap)) | |
1308 | continue; | |
2c8cdd6e | 1309 | |
36920044 QW |
1310 | if (stripe < rbio->nr_data) { |
1311 | sector = sector_in_rbio(rbio, stripe, sectornr, 1); | |
1312 | if (!sector) | |
1313 | continue; | |
1314 | } else { | |
1315 | sector = rbio_stripe_sector(rbio, stripe, sectornr); | |
2c8cdd6e | 1316 | } |
36920044 | 1317 | |
6486d21c | 1318 | ret = rbio_add_io_sector(rbio, bio_list, sector, |
36920044 | 1319 | rbio->bioc->tgtdev_map[stripe], |
ff18a4af | 1320 | sectornr, REQ_OP_WRITE); |
36920044 | 1321 | if (ret) |
6486d21c | 1322 | goto error; |
2c8cdd6e MX |
1323 | } |
1324 | ||
6486d21c QW |
1325 | return 0; |
1326 | error: | |
801fcfc5 | 1327 | bio_list_put(bio_list); |
6486d21c QW |
1328 | return -EIO; |
1329 | } | |
1330 | ||
2942a50d QW |
1331 | static void set_rbio_range_error(struct btrfs_raid_bio *rbio, struct bio *bio) |
1332 | { | |
1333 | struct btrfs_fs_info *fs_info = rbio->bioc->fs_info; | |
1334 | u32 offset = (bio->bi_iter.bi_sector << SECTOR_SHIFT) - | |
1335 | rbio->bioc->raid_map[0]; | |
1336 | int total_nr_sector = offset >> fs_info->sectorsize_bits; | |
1337 | ||
1338 | ASSERT(total_nr_sector < rbio->nr_data * rbio->stripe_nsectors); | |
1339 | ||
1340 | bitmap_set(rbio->error_bitmap, total_nr_sector, | |
1341 | bio->bi_iter.bi_size >> fs_info->sectorsize_bits); | |
1342 | ||
1343 | /* | |
1344 | * Special handling for raid56_alloc_missing_rbio() used by | |
1345 | * scrub/replace. Unlike call path in raid56_parity_recover(), they | |
1346 | * pass an empty bio here. Thus we have to find out the missing device | |
1347 | * and mark the stripe error instead. | |
1348 | */ | |
1349 | if (bio->bi_iter.bi_size == 0) { | |
1350 | bool found_missing = false; | |
1351 | int stripe_nr; | |
1352 | ||
1353 | for (stripe_nr = 0; stripe_nr < rbio->real_stripes; stripe_nr++) { | |
1354 | if (!rbio->bioc->stripes[stripe_nr].dev->bdev) { | |
1355 | found_missing = true; | |
1356 | bitmap_set(rbio->error_bitmap, | |
1357 | stripe_nr * rbio->stripe_nsectors, | |
1358 | rbio->stripe_nsectors); | |
1359 | } | |
1360 | } | |
1361 | ASSERT(found_missing); | |
1362 | } | |
1363 | } | |
1364 | ||
5fdb7afc | 1365 | /* |
67da05b3 | 1366 | * For subpage case, we can no longer set page Up-to-date directly for |
5fdb7afc QW |
1367 | * stripe_pages[], thus we need to locate the sector. |
1368 | */ | |
1369 | static struct sector_ptr *find_stripe_sector(struct btrfs_raid_bio *rbio, | |
1370 | struct page *page, | |
1371 | unsigned int pgoff) | |
1372 | { | |
1373 | int i; | |
1374 | ||
1375 | for (i = 0; i < rbio->nr_sectors; i++) { | |
1376 | struct sector_ptr *sector = &rbio->stripe_sectors[i]; | |
1377 | ||
1378 | if (sector->page == page && sector->pgoff == pgoff) | |
1379 | return sector; | |
1380 | } | |
1381 | return NULL; | |
1382 | } | |
1383 | ||
53b381b3 DW |
1384 | /* |
1385 | * this sets each page in the bio uptodate. It should only be used on private | |
1386 | * rbio pages, nothing that comes in from the higher layers | |
1387 | */ | |
5fdb7afc | 1388 | static void set_bio_pages_uptodate(struct btrfs_raid_bio *rbio, struct bio *bio) |
53b381b3 | 1389 | { |
5fdb7afc | 1390 | const u32 sectorsize = rbio->bioc->fs_info->sectorsize; |
0198e5b7 | 1391 | struct bio_vec *bvec; |
6dc4f100 | 1392 | struct bvec_iter_all iter_all; |
6592e58c | 1393 | |
0198e5b7 | 1394 | ASSERT(!bio_flagged(bio, BIO_CLONED)); |
53b381b3 | 1395 | |
5fdb7afc QW |
1396 | bio_for_each_segment_all(bvec, bio, iter_all) { |
1397 | struct sector_ptr *sector; | |
1398 | int pgoff; | |
1399 | ||
1400 | for (pgoff = bvec->bv_offset; pgoff - bvec->bv_offset < bvec->bv_len; | |
1401 | pgoff += sectorsize) { | |
1402 | sector = find_stripe_sector(rbio, bvec->bv_page, pgoff); | |
1403 | ASSERT(sector); | |
1404 | if (sector) | |
1405 | sector->uptodate = 1; | |
1406 | } | |
1407 | } | |
53b381b3 DW |
1408 | } |
1409 | ||
2942a50d QW |
1410 | static int get_bio_sector_nr(struct btrfs_raid_bio *rbio, struct bio *bio) |
1411 | { | |
1412 | struct bio_vec *bv = bio_first_bvec_all(bio); | |
1413 | int i; | |
1414 | ||
1415 | for (i = 0; i < rbio->nr_sectors; i++) { | |
1416 | struct sector_ptr *sector; | |
1417 | ||
1418 | sector = &rbio->stripe_sectors[i]; | |
1419 | if (sector->page == bv->bv_page && sector->pgoff == bv->bv_offset) | |
1420 | break; | |
1421 | sector = &rbio->bio_sectors[i]; | |
1422 | if (sector->page == bv->bv_page && sector->pgoff == bv->bv_offset) | |
1423 | break; | |
1424 | } | |
1425 | ASSERT(i < rbio->nr_sectors); | |
1426 | return i; | |
1427 | } | |
1428 | ||
1429 | static void rbio_update_error_bitmap(struct btrfs_raid_bio *rbio, struct bio *bio) | |
1430 | { | |
1431 | int total_sector_nr = get_bio_sector_nr(rbio, bio); | |
1432 | u32 bio_size = 0; | |
1433 | struct bio_vec *bvec; | |
a9ad4d87 | 1434 | int i; |
2942a50d | 1435 | |
c9a43aaf | 1436 | bio_for_each_bvec_all(bvec, bio, i) |
2942a50d QW |
1437 | bio_size += bvec->bv_len; |
1438 | ||
a9ad4d87 QW |
1439 | /* |
1440 | * Since we can have multiple bios touching the error_bitmap, we cannot | |
1441 | * call bitmap_set() without protection. | |
1442 | * | |
1443 | * Instead use set_bit() for each bit, as set_bit() itself is atomic. | |
1444 | */ | |
1445 | for (i = total_sector_nr; i < total_sector_nr + | |
1446 | (bio_size >> rbio->bioc->fs_info->sectorsize_bits); i++) | |
1447 | set_bit(i, rbio->error_bitmap); | |
2942a50d QW |
1448 | } |
1449 | ||
7a315072 QW |
1450 | /* Verify the data sectors at read time. */ |
1451 | static void verify_bio_data_sectors(struct btrfs_raid_bio *rbio, | |
1452 | struct bio *bio) | |
1453 | { | |
1454 | struct btrfs_fs_info *fs_info = rbio->bioc->fs_info; | |
1455 | int total_sector_nr = get_bio_sector_nr(rbio, bio); | |
1456 | struct bio_vec *bvec; | |
1457 | struct bvec_iter_all iter_all; | |
1458 | ||
1459 | /* No data csum for the whole stripe, no need to verify. */ | |
1460 | if (!rbio->csum_bitmap || !rbio->csum_buf) | |
1461 | return; | |
1462 | ||
1463 | /* P/Q stripes, they have no data csum to verify against. */ | |
1464 | if (total_sector_nr >= rbio->nr_data * rbio->stripe_nsectors) | |
1465 | return; | |
1466 | ||
1467 | bio_for_each_segment_all(bvec, bio, iter_all) { | |
1468 | int bv_offset; | |
1469 | ||
1470 | for (bv_offset = bvec->bv_offset; | |
1471 | bv_offset < bvec->bv_offset + bvec->bv_len; | |
1472 | bv_offset += fs_info->sectorsize, total_sector_nr++) { | |
1473 | u8 csum_buf[BTRFS_CSUM_SIZE]; | |
1474 | u8 *expected_csum = rbio->csum_buf + | |
1475 | total_sector_nr * fs_info->csum_size; | |
1476 | int ret; | |
1477 | ||
1478 | /* No csum for this sector, skip to the next sector. */ | |
1479 | if (!test_bit(total_sector_nr, rbio->csum_bitmap)) | |
1480 | continue; | |
1481 | ||
1482 | ret = btrfs_check_sector_csum(fs_info, bvec->bv_page, | |
1483 | bv_offset, csum_buf, expected_csum); | |
1484 | if (ret < 0) | |
1485 | set_bit(total_sector_nr, rbio->error_bitmap); | |
1486 | } | |
1487 | } | |
1488 | } | |
1489 | ||
d817ce35 QW |
1490 | static void raid_wait_read_end_io(struct bio *bio) |
1491 | { | |
1492 | struct btrfs_raid_bio *rbio = bio->bi_private; | |
1493 | ||
7a315072 | 1494 | if (bio->bi_status) { |
2942a50d | 1495 | rbio_update_error_bitmap(rbio, bio); |
7a315072 | 1496 | } else { |
d817ce35 | 1497 | set_bio_pages_uptodate(rbio, bio); |
7a315072 QW |
1498 | verify_bio_data_sectors(rbio, bio); |
1499 | } | |
d817ce35 QW |
1500 | |
1501 | bio_put(bio); | |
1502 | if (atomic_dec_and_test(&rbio->stripes_pending)) | |
1503 | wake_up(&rbio->io_wait); | |
1504 | } | |
1505 | ||
1c76fb7b | 1506 | static void submit_read_wait_bio_list(struct btrfs_raid_bio *rbio, |
d817ce35 QW |
1507 | struct bio_list *bio_list) |
1508 | { | |
1509 | struct bio *bio; | |
1510 | ||
1511 | atomic_set(&rbio->stripes_pending, bio_list_size(bio_list)); | |
1512 | while ((bio = bio_list_pop(bio_list))) { | |
1513 | bio->bi_end_io = raid_wait_read_end_io; | |
1514 | ||
1515 | if (trace_raid56_scrub_read_recover_enabled()) { | |
1516 | struct raid56_bio_trace_info trace_info = { 0 }; | |
1517 | ||
1518 | bio_get_trace_info(rbio, bio, &trace_info); | |
1519 | trace_raid56_scrub_read_recover(rbio, bio, &trace_info); | |
1520 | } | |
1521 | submit_bio(bio); | |
1522 | } | |
1c76fb7b CH |
1523 | |
1524 | wait_event(rbio->io_wait, atomic_read(&rbio->stripes_pending) == 0); | |
d817ce35 QW |
1525 | } |
1526 | ||
5eb30ee2 QW |
1527 | static int alloc_rbio_data_pages(struct btrfs_raid_bio *rbio) |
1528 | { | |
1529 | const int data_pages = rbio->nr_data * rbio->stripe_npages; | |
1530 | int ret; | |
1531 | ||
1532 | ret = btrfs_alloc_page_array(data_pages, rbio->stripe_pages); | |
1533 | if (ret < 0) | |
1534 | return ret; | |
1535 | ||
1536 | index_stripe_sectors(rbio); | |
1537 | return 0; | |
1538 | } | |
1539 | ||
6ac0f488 CM |
1540 | /* |
1541 | * We use plugging call backs to collect full stripes. | |
1542 | * Any time we get a partial stripe write while plugged | |
1543 | * we collect it into a list. When the unplug comes down, | |
1544 | * we sort the list by logical block number and merge | |
1545 | * everything we can into the same rbios | |
1546 | */ | |
1547 | struct btrfs_plug_cb { | |
1548 | struct blk_plug_cb cb; | |
1549 | struct btrfs_fs_info *info; | |
1550 | struct list_head rbio_list; | |
385de0ef | 1551 | struct work_struct work; |
6ac0f488 CM |
1552 | }; |
1553 | ||
1554 | /* | |
1555 | * rbios on the plug list are sorted for easier merging. | |
1556 | */ | |
4f0f586b ST |
1557 | static int plug_cmp(void *priv, const struct list_head *a, |
1558 | const struct list_head *b) | |
6ac0f488 | 1559 | { |
214cc184 DS |
1560 | const struct btrfs_raid_bio *ra = container_of(a, struct btrfs_raid_bio, |
1561 | plug_list); | |
1562 | const struct btrfs_raid_bio *rb = container_of(b, struct btrfs_raid_bio, | |
1563 | plug_list); | |
4f024f37 KO |
1564 | u64 a_sector = ra->bio_list.head->bi_iter.bi_sector; |
1565 | u64 b_sector = rb->bio_list.head->bi_iter.bi_sector; | |
6ac0f488 CM |
1566 | |
1567 | if (a_sector < b_sector) | |
1568 | return -1; | |
1569 | if (a_sector > b_sector) | |
1570 | return 1; | |
1571 | return 0; | |
1572 | } | |
1573 | ||
93723095 | 1574 | static void raid_unplug(struct blk_plug_cb *cb, bool from_schedule) |
6ac0f488 | 1575 | { |
93723095 | 1576 | struct btrfs_plug_cb *plug = container_of(cb, struct btrfs_plug_cb, cb); |
6ac0f488 CM |
1577 | struct btrfs_raid_bio *cur; |
1578 | struct btrfs_raid_bio *last = NULL; | |
1579 | ||
6ac0f488 | 1580 | list_sort(NULL, &plug->rbio_list, plug_cmp); |
93723095 | 1581 | |
6ac0f488 CM |
1582 | while (!list_empty(&plug->rbio_list)) { |
1583 | cur = list_entry(plug->rbio_list.next, | |
1584 | struct btrfs_raid_bio, plug_list); | |
1585 | list_del_init(&cur->plug_list); | |
1586 | ||
1587 | if (rbio_is_full(cur)) { | |
93723095 QW |
1588 | /* We have a full stripe, queue it down. */ |
1589 | start_async_work(cur, rmw_rbio_work); | |
6ac0f488 CM |
1590 | continue; |
1591 | } | |
1592 | if (last) { | |
1593 | if (rbio_can_merge(last, cur)) { | |
1594 | merge_rbio(last, cur); | |
ff2b64a2 | 1595 | free_raid_bio(cur); |
6ac0f488 | 1596 | continue; |
6ac0f488 | 1597 | } |
93723095 | 1598 | start_async_work(last, rmw_rbio_work); |
6ac0f488 CM |
1599 | } |
1600 | last = cur; | |
1601 | } | |
93723095 QW |
1602 | if (last) |
1603 | start_async_work(last, rmw_rbio_work); | |
6ac0f488 CM |
1604 | kfree(plug); |
1605 | } | |
1606 | ||
bd8f7e62 QW |
1607 | /* Add the original bio into rbio->bio_list, and update rbio::dbitmap. */ |
1608 | static void rbio_add_bio(struct btrfs_raid_bio *rbio, struct bio *orig_bio) | |
1609 | { | |
1610 | const struct btrfs_fs_info *fs_info = rbio->bioc->fs_info; | |
1611 | const u64 orig_logical = orig_bio->bi_iter.bi_sector << SECTOR_SHIFT; | |
1612 | const u64 full_stripe_start = rbio->bioc->raid_map[0]; | |
1613 | const u32 orig_len = orig_bio->bi_iter.bi_size; | |
1614 | const u32 sectorsize = fs_info->sectorsize; | |
1615 | u64 cur_logical; | |
1616 | ||
1617 | ASSERT(orig_logical >= full_stripe_start && | |
1618 | orig_logical + orig_len <= full_stripe_start + | |
ff18a4af | 1619 | rbio->nr_data * BTRFS_STRIPE_LEN); |
bd8f7e62 QW |
1620 | |
1621 | bio_list_add(&rbio->bio_list, orig_bio); | |
1622 | rbio->bio_list_bytes += orig_bio->bi_iter.bi_size; | |
1623 | ||
1624 | /* Update the dbitmap. */ | |
1625 | for (cur_logical = orig_logical; cur_logical < orig_logical + orig_len; | |
1626 | cur_logical += sectorsize) { | |
1627 | int bit = ((u32)(cur_logical - full_stripe_start) >> | |
1628 | fs_info->sectorsize_bits) % rbio->stripe_nsectors; | |
1629 | ||
1630 | set_bit(bit, &rbio->dbitmap); | |
1631 | } | |
1632 | } | |
1633 | ||
53b381b3 DW |
1634 | /* |
1635 | * our main entry point for writes from the rest of the FS. | |
1636 | */ | |
31683f4a | 1637 | void raid56_parity_write(struct bio *bio, struct btrfs_io_context *bioc) |
53b381b3 | 1638 | { |
6a258d72 | 1639 | struct btrfs_fs_info *fs_info = bioc->fs_info; |
53b381b3 | 1640 | struct btrfs_raid_bio *rbio; |
6ac0f488 CM |
1641 | struct btrfs_plug_cb *plug = NULL; |
1642 | struct blk_plug_cb *cb; | |
53b381b3 | 1643 | |
ff18a4af | 1644 | rbio = alloc_rbio(fs_info, bioc); |
af8e2d1d | 1645 | if (IS_ERR(rbio)) { |
abb49e87 CH |
1646 | bio->bi_status = errno_to_blk_status(PTR_ERR(rbio)); |
1647 | bio_endio(bio); | |
1648 | return; | |
af8e2d1d | 1649 | } |
1b94b556 | 1650 | rbio->operation = BTRFS_RBIO_WRITE; |
bd8f7e62 | 1651 | rbio_add_bio(rbio, bio); |
6ac0f488 CM |
1652 | |
1653 | /* | |
93723095 | 1654 | * Don't plug on full rbios, just get them out the door |
6ac0f488 CM |
1655 | * as quickly as we can |
1656 | */ | |
abb49e87 CH |
1657 | if (!rbio_is_full(rbio)) { |
1658 | cb = blk_check_plugged(raid_unplug, fs_info, sizeof(*plug)); | |
1659 | if (cb) { | |
1660 | plug = container_of(cb, struct btrfs_plug_cb, cb); | |
1661 | if (!plug->info) { | |
1662 | plug->info = fs_info; | |
1663 | INIT_LIST_HEAD(&plug->rbio_list); | |
1664 | } | |
1665 | list_add_tail(&rbio->plug_list, &plug->rbio_list); | |
1666 | return; | |
6ac0f488 | 1667 | } |
6ac0f488 | 1668 | } |
abb49e87 | 1669 | |
93723095 QW |
1670 | /* |
1671 | * Either we don't have any existing plug, or we're doing a full stripe, | |
abb49e87 | 1672 | * queue the rmw work now. |
93723095 QW |
1673 | */ |
1674 | start_async_work(rbio, rmw_rbio_work); | |
53b381b3 DW |
1675 | } |
1676 | ||
7a315072 QW |
1677 | static int verify_one_sector(struct btrfs_raid_bio *rbio, |
1678 | int stripe_nr, int sector_nr) | |
1679 | { | |
1680 | struct btrfs_fs_info *fs_info = rbio->bioc->fs_info; | |
1681 | struct sector_ptr *sector; | |
1682 | u8 csum_buf[BTRFS_CSUM_SIZE]; | |
1683 | u8 *csum_expected; | |
1684 | int ret; | |
1685 | ||
1686 | if (!rbio->csum_bitmap || !rbio->csum_buf) | |
1687 | return 0; | |
1688 | ||
1689 | /* No way to verify P/Q as they are not covered by data csum. */ | |
1690 | if (stripe_nr >= rbio->nr_data) | |
1691 | return 0; | |
1692 | /* | |
1693 | * If we're rebuilding a read, we have to use pages from the | |
1694 | * bio list if possible. | |
1695 | */ | |
1696 | if ((rbio->operation == BTRFS_RBIO_READ_REBUILD || | |
1697 | rbio->operation == BTRFS_RBIO_REBUILD_MISSING)) { | |
1698 | sector = sector_in_rbio(rbio, stripe_nr, sector_nr, 0); | |
1699 | } else { | |
1700 | sector = rbio_stripe_sector(rbio, stripe_nr, sector_nr); | |
1701 | } | |
1702 | ||
1703 | ASSERT(sector->page); | |
1704 | ||
1705 | csum_expected = rbio->csum_buf + | |
1706 | (stripe_nr * rbio->stripe_nsectors + sector_nr) * | |
1707 | fs_info->csum_size; | |
1708 | ret = btrfs_check_sector_csum(fs_info, sector->page, sector->pgoff, | |
1709 | csum_buf, csum_expected); | |
1710 | return ret; | |
1711 | } | |
1712 | ||
9c5ff9b4 QW |
1713 | /* |
1714 | * Recover a vertical stripe specified by @sector_nr. | |
1715 | * @*pointers are the pre-allocated pointers by the caller, so we don't | |
1716 | * need to allocate/free the pointers again and again. | |
1717 | */ | |
75b47033 QW |
1718 | static int recover_vertical(struct btrfs_raid_bio *rbio, int sector_nr, |
1719 | void **pointers, void **unmap_array) | |
9c5ff9b4 QW |
1720 | { |
1721 | struct btrfs_fs_info *fs_info = rbio->bioc->fs_info; | |
1722 | struct sector_ptr *sector; | |
1723 | const u32 sectorsize = fs_info->sectorsize; | |
75b47033 QW |
1724 | int found_errors; |
1725 | int faila; | |
1726 | int failb; | |
9c5ff9b4 | 1727 | int stripe_nr; |
7a315072 | 1728 | int ret = 0; |
9c5ff9b4 QW |
1729 | |
1730 | /* | |
1731 | * Now we just use bitmap to mark the horizontal stripes in | |
1732 | * which we have data when doing parity scrub. | |
1733 | */ | |
1734 | if (rbio->operation == BTRFS_RBIO_PARITY_SCRUB && | |
1735 | !test_bit(sector_nr, &rbio->dbitmap)) | |
75b47033 QW |
1736 | return 0; |
1737 | ||
1738 | found_errors = get_rbio_veritical_errors(rbio, sector_nr, &faila, | |
1739 | &failb); | |
1740 | /* | |
67da05b3 | 1741 | * No errors in the vertical stripe, skip it. Can happen for recovery |
75b47033 QW |
1742 | * which only part of a stripe failed csum check. |
1743 | */ | |
1744 | if (!found_errors) | |
1745 | return 0; | |
1746 | ||
1747 | if (found_errors > rbio->bioc->max_errors) | |
1748 | return -EIO; | |
9c5ff9b4 QW |
1749 | |
1750 | /* | |
1751 | * Setup our array of pointers with sectors from each stripe | |
1752 | * | |
1753 | * NOTE: store a duplicate array of pointers to preserve the | |
1754 | * pointer order. | |
1755 | */ | |
1756 | for (stripe_nr = 0; stripe_nr < rbio->real_stripes; stripe_nr++) { | |
1757 | /* | |
75b47033 QW |
1758 | * If we're rebuilding a read, we have to use pages from the |
1759 | * bio list if possible. | |
9c5ff9b4 QW |
1760 | */ |
1761 | if ((rbio->operation == BTRFS_RBIO_READ_REBUILD || | |
75b47033 | 1762 | rbio->operation == BTRFS_RBIO_REBUILD_MISSING)) { |
9c5ff9b4 QW |
1763 | sector = sector_in_rbio(rbio, stripe_nr, sector_nr, 0); |
1764 | } else { | |
1765 | sector = rbio_stripe_sector(rbio, stripe_nr, sector_nr); | |
1766 | } | |
1767 | ASSERT(sector->page); | |
1768 | pointers[stripe_nr] = kmap_local_page(sector->page) + | |
1769 | sector->pgoff; | |
1770 | unmap_array[stripe_nr] = pointers[stripe_nr]; | |
1771 | } | |
1772 | ||
1773 | /* All raid6 handling here */ | |
1774 | if (rbio->bioc->map_type & BTRFS_BLOCK_GROUP_RAID6) { | |
1775 | /* Single failure, rebuild from parity raid5 style */ | |
1776 | if (failb < 0) { | |
1777 | if (faila == rbio->nr_data) | |
1778 | /* | |
1779 | * Just the P stripe has failed, without | |
1780 | * a bad data or Q stripe. | |
1781 | * We have nothing to do, just skip the | |
1782 | * recovery for this stripe. | |
1783 | */ | |
1784 | goto cleanup; | |
1785 | /* | |
1786 | * a single failure in raid6 is rebuilt | |
1787 | * in the pstripe code below | |
1788 | */ | |
1789 | goto pstripe; | |
1790 | } | |
1791 | ||
1792 | /* | |
1793 | * If the q stripe is failed, do a pstripe reconstruction from | |
1794 | * the xors. | |
1795 | * If both the q stripe and the P stripe are failed, we're | |
1796 | * here due to a crc mismatch and we can't give them the | |
1797 | * data they want. | |
1798 | */ | |
1799 | if (rbio->bioc->raid_map[failb] == RAID6_Q_STRIPE) { | |
1800 | if (rbio->bioc->raid_map[faila] == | |
1801 | RAID5_P_STRIPE) | |
1802 | /* | |
1803 | * Only P and Q are corrupted. | |
1804 | * We only care about data stripes recovery, | |
1805 | * can skip this vertical stripe. | |
1806 | */ | |
1807 | goto cleanup; | |
1808 | /* | |
1809 | * Otherwise we have one bad data stripe and | |
1810 | * a good P stripe. raid5! | |
1811 | */ | |
1812 | goto pstripe; | |
1813 | } | |
1814 | ||
1815 | if (rbio->bioc->raid_map[failb] == RAID5_P_STRIPE) { | |
1816 | raid6_datap_recov(rbio->real_stripes, sectorsize, | |
1817 | faila, pointers); | |
1818 | } else { | |
1819 | raid6_2data_recov(rbio->real_stripes, sectorsize, | |
1820 | faila, failb, pointers); | |
1821 | } | |
1822 | } else { | |
1823 | void *p; | |
1824 | ||
1825 | /* Rebuild from P stripe here (raid5 or raid6). */ | |
1826 | ASSERT(failb == -1); | |
1827 | pstripe: | |
1828 | /* Copy parity block into failed block to start with */ | |
1829 | memcpy(pointers[faila], pointers[rbio->nr_data], sectorsize); | |
1830 | ||
1831 | /* Rearrange the pointer array */ | |
1832 | p = pointers[faila]; | |
1833 | for (stripe_nr = faila; stripe_nr < rbio->nr_data - 1; | |
1834 | stripe_nr++) | |
1835 | pointers[stripe_nr] = pointers[stripe_nr + 1]; | |
1836 | pointers[rbio->nr_data - 1] = p; | |
1837 | ||
1838 | /* Xor in the rest */ | |
1839 | run_xor(pointers, rbio->nr_data - 1, sectorsize); | |
1840 | ||
1841 | } | |
1842 | ||
1843 | /* | |
1844 | * No matter if this is a RMW or recovery, we should have all | |
1845 | * failed sectors repaired in the vertical stripe, thus they are now | |
1846 | * uptodate. | |
1847 | * Especially if we determine to cache the rbio, we need to | |
1848 | * have at least all data sectors uptodate. | |
7a315072 QW |
1849 | * |
1850 | * If possible, also check if the repaired sector matches its data | |
1851 | * checksum. | |
9c5ff9b4 | 1852 | */ |
75b47033 | 1853 | if (faila >= 0) { |
7a315072 QW |
1854 | ret = verify_one_sector(rbio, faila, sector_nr); |
1855 | if (ret < 0) | |
1856 | goto cleanup; | |
1857 | ||
75b47033 | 1858 | sector = rbio_stripe_sector(rbio, faila, sector_nr); |
9c5ff9b4 QW |
1859 | sector->uptodate = 1; |
1860 | } | |
75b47033 | 1861 | if (failb >= 0) { |
f7c11aff | 1862 | ret = verify_one_sector(rbio, failb, sector_nr); |
7a315072 QW |
1863 | if (ret < 0) |
1864 | goto cleanup; | |
1865 | ||
75b47033 | 1866 | sector = rbio_stripe_sector(rbio, failb, sector_nr); |
9c5ff9b4 QW |
1867 | sector->uptodate = 1; |
1868 | } | |
1869 | ||
1870 | cleanup: | |
1871 | for (stripe_nr = rbio->real_stripes - 1; stripe_nr >= 0; stripe_nr--) | |
1872 | kunmap_local(unmap_array[stripe_nr]); | |
7a315072 | 1873 | return ret; |
9c5ff9b4 QW |
1874 | } |
1875 | ||
ec936b03 | 1876 | static int recover_sectors(struct btrfs_raid_bio *rbio) |
53b381b3 | 1877 | { |
9c5ff9b4 QW |
1878 | void **pointers = NULL; |
1879 | void **unmap_array = NULL; | |
ec936b03 QW |
1880 | int sectornr; |
1881 | int ret = 0; | |
53b381b3 | 1882 | |
07e4d380 | 1883 | /* |
ec936b03 QW |
1884 | * @pointers array stores the pointer for each sector. |
1885 | * | |
1886 | * @unmap_array stores copy of pointers that does not get reordered | |
1887 | * during reconstruction so that kunmap_local works. | |
07e4d380 | 1888 | */ |
31e818fe | 1889 | pointers = kcalloc(rbio->real_stripes, sizeof(void *), GFP_NOFS); |
94a0b58d | 1890 | unmap_array = kcalloc(rbio->real_stripes, sizeof(void *), GFP_NOFS); |
ec936b03 QW |
1891 | if (!pointers || !unmap_array) { |
1892 | ret = -ENOMEM; | |
1893 | goto out; | |
94a0b58d IW |
1894 | } |
1895 | ||
b4ee1782 OS |
1896 | if (rbio->operation == BTRFS_RBIO_READ_REBUILD || |
1897 | rbio->operation == BTRFS_RBIO_REBUILD_MISSING) { | |
53b381b3 DW |
1898 | spin_lock_irq(&rbio->bio_list_lock); |
1899 | set_bit(RBIO_RMW_LOCKED_BIT, &rbio->flags); | |
1900 | spin_unlock_irq(&rbio->bio_list_lock); | |
1901 | } | |
1902 | ||
1903 | index_rbio_pages(rbio); | |
1904 | ||
75b47033 QW |
1905 | for (sectornr = 0; sectornr < rbio->stripe_nsectors; sectornr++) { |
1906 | ret = recover_vertical(rbio, sectornr, pointers, unmap_array); | |
1907 | if (ret < 0) | |
1908 | break; | |
1909 | } | |
53b381b3 | 1910 | |
ec936b03 | 1911 | out: |
53b381b3 | 1912 | kfree(pointers); |
ec936b03 QW |
1913 | kfree(unmap_array); |
1914 | return ret; | |
1915 | } | |
1916 | ||
d838d05e | 1917 | static int recover_rbio(struct btrfs_raid_bio *rbio) |
53b381b3 | 1918 | { |
d838d05e | 1919 | struct bio_list bio_list = BIO_EMPTY_LIST; |
d31968d9 QW |
1920 | int total_sector_nr; |
1921 | int ret = 0; | |
53b381b3 | 1922 | |
d838d05e CH |
1923 | /* |
1924 | * Either we're doing recover for a read failure or degraded write, | |
1925 | * caller should have set error bitmap correctly. | |
1926 | */ | |
1927 | ASSERT(bitmap_weight(rbio->error_bitmap, rbio->nr_sectors)); | |
1928 | ||
1929 | /* For recovery, we need to read all sectors including P/Q. */ | |
1930 | ret = alloc_rbio_pages(rbio); | |
1931 | if (ret < 0) | |
1932 | return ret; | |
1933 | ||
1934 | index_rbio_pages(rbio); | |
1935 | ||
53b381b3 | 1936 | /* |
f6065f8e QW |
1937 | * Read everything that hasn't failed. However this time we will |
1938 | * not trust any cached sector. | |
1939 | * As we may read out some stale data but higher layer is not reading | |
1940 | * that stale part. | |
1941 | * | |
1942 | * So here we always re-read everything in recovery path. | |
53b381b3 | 1943 | */ |
ef340fcc QW |
1944 | for (total_sector_nr = 0; total_sector_nr < rbio->nr_sectors; |
1945 | total_sector_nr++) { | |
1946 | int stripe = total_sector_nr / rbio->stripe_nsectors; | |
1947 | int sectornr = total_sector_nr % rbio->stripe_nsectors; | |
1948 | struct sector_ptr *sector; | |
1949 | ||
75b47033 QW |
1950 | /* |
1951 | * Skip the range which has error. It can be a range which is | |
1952 | * marked error (for csum mismatch), or it can be a missing | |
1953 | * device. | |
1954 | */ | |
1955 | if (!rbio->bioc->stripes[stripe].dev->bdev || | |
1956 | test_bit(total_sector_nr, rbio->error_bitmap)) { | |
1957 | /* | |
1958 | * Also set the error bit for missing device, which | |
1959 | * may not yet have its error bit set. | |
1960 | */ | |
1961 | set_bit(total_sector_nr, rbio->error_bitmap); | |
53b381b3 | 1962 | continue; |
5588383e | 1963 | } |
75b47033 | 1964 | |
ef340fcc | 1965 | sector = rbio_stripe_sector(rbio, stripe, sectornr); |
d838d05e | 1966 | ret = rbio_add_io_sector(rbio, &bio_list, sector, stripe, |
ff18a4af | 1967 | sectornr, REQ_OP_READ); |
d838d05e CH |
1968 | if (ret < 0) { |
1969 | bio_list_put(&bio_list); | |
1970 | return ret; | |
1971 | } | |
53b381b3 | 1972 | } |
d817ce35 | 1973 | |
1c76fb7b | 1974 | submit_read_wait_bio_list(rbio, &bio_list); |
d838d05e | 1975 | return recover_sectors(rbio); |
d817ce35 QW |
1976 | } |
1977 | ||
1978 | static void recover_rbio_work(struct work_struct *work) | |
1979 | { | |
1980 | struct btrfs_raid_bio *rbio; | |
1981 | int ret; | |
1982 | ||
1983 | rbio = container_of(work, struct btrfs_raid_bio, work); | |
1984 | ||
1985 | ret = lock_stripe_add(rbio); | |
1986 | if (ret == 0) { | |
1987 | ret = recover_rbio(rbio); | |
1988 | rbio_orig_end_io(rbio, errno_to_blk_status(ret)); | |
1989 | } | |
1990 | } | |
1991 | ||
1992 | static void recover_rbio_work_locked(struct work_struct *work) | |
1993 | { | |
1994 | struct btrfs_raid_bio *rbio; | |
1995 | int ret; | |
1996 | ||
1997 | rbio = container_of(work, struct btrfs_raid_bio, work); | |
1998 | ||
1999 | ret = recover_rbio(rbio); | |
2000 | rbio_orig_end_io(rbio, errno_to_blk_status(ret)); | |
2001 | } | |
2002 | ||
75b47033 QW |
2003 | static void set_rbio_raid6_extra_error(struct btrfs_raid_bio *rbio, int mirror_num) |
2004 | { | |
2005 | bool found = false; | |
2006 | int sector_nr; | |
2007 | ||
2008 | /* | |
2009 | * This is for RAID6 extra recovery tries, thus mirror number should | |
2010 | * be large than 2. | |
2011 | * Mirror 1 means read from data stripes. Mirror 2 means rebuild using | |
2012 | * RAID5 methods. | |
2013 | */ | |
2014 | ASSERT(mirror_num > 2); | |
2015 | for (sector_nr = 0; sector_nr < rbio->stripe_nsectors; sector_nr++) { | |
2016 | int found_errors; | |
2017 | int faila; | |
2018 | int failb; | |
2019 | ||
2020 | found_errors = get_rbio_veritical_errors(rbio, sector_nr, | |
2021 | &faila, &failb); | |
2022 | /* This vertical stripe doesn't have errors. */ | |
2023 | if (!found_errors) | |
2024 | continue; | |
2025 | ||
2026 | /* | |
2027 | * If we found errors, there should be only one error marked | |
2028 | * by previous set_rbio_range_error(). | |
2029 | */ | |
2030 | ASSERT(found_errors == 1); | |
2031 | found = true; | |
2032 | ||
2033 | /* Now select another stripe to mark as error. */ | |
2034 | failb = rbio->real_stripes - (mirror_num - 1); | |
2035 | if (failb <= faila) | |
2036 | failb--; | |
2037 | ||
2038 | /* Set the extra bit in error bitmap. */ | |
2039 | if (failb >= 0) | |
2040 | set_bit(failb * rbio->stripe_nsectors + sector_nr, | |
2041 | rbio->error_bitmap); | |
2042 | } | |
2043 | ||
2044 | /* We should found at least one vertical stripe with error.*/ | |
2045 | ASSERT(found); | |
2046 | } | |
2047 | ||
53b381b3 DW |
2048 | /* |
2049 | * the main entry point for reads from the higher layers. This | |
2050 | * is really only called when the normal read path had a failure, | |
2051 | * so we assume the bio they send down corresponds to a failed part | |
2052 | * of the drive. | |
2053 | */ | |
6065fd95 | 2054 | void raid56_parity_recover(struct bio *bio, struct btrfs_io_context *bioc, |
f1c29379 | 2055 | int mirror_num) |
53b381b3 | 2056 | { |
6a258d72 | 2057 | struct btrfs_fs_info *fs_info = bioc->fs_info; |
53b381b3 | 2058 | struct btrfs_raid_bio *rbio; |
53b381b3 | 2059 | |
ff18a4af | 2060 | rbio = alloc_rbio(fs_info, bioc); |
af8e2d1d | 2061 | if (IS_ERR(rbio)) { |
6065fd95 | 2062 | bio->bi_status = errno_to_blk_status(PTR_ERR(rbio)); |
d817ce35 QW |
2063 | bio_endio(bio); |
2064 | return; | |
af8e2d1d | 2065 | } |
53b381b3 | 2066 | |
1b94b556 | 2067 | rbio->operation = BTRFS_RBIO_READ_REBUILD; |
bd8f7e62 | 2068 | rbio_add_bio(rbio, bio); |
53b381b3 | 2069 | |
2942a50d QW |
2070 | set_rbio_range_error(rbio, bio); |
2071 | ||
53b381b3 | 2072 | /* |
8810f751 LB |
2073 | * Loop retry: |
2074 | * for 'mirror == 2', reconstruct from all other stripes. | |
2075 | * for 'mirror_num > 2', select a stripe to fail on every retry. | |
53b381b3 | 2076 | */ |
ad3daf1c | 2077 | if (mirror_num > 2) |
75b47033 | 2078 | set_rbio_raid6_extra_error(rbio, mirror_num); |
53b381b3 | 2079 | |
d817ce35 | 2080 | start_async_work(rbio, recover_rbio_work); |
53b381b3 DW |
2081 | } |
2082 | ||
c5a41562 QW |
2083 | static void fill_data_csums(struct btrfs_raid_bio *rbio) |
2084 | { | |
2085 | struct btrfs_fs_info *fs_info = rbio->bioc->fs_info; | |
2086 | struct btrfs_root *csum_root = btrfs_csum_root(fs_info, | |
2087 | rbio->bioc->raid_map[0]); | |
2088 | const u64 start = rbio->bioc->raid_map[0]; | |
2089 | const u32 len = (rbio->nr_data * rbio->stripe_nsectors) << | |
2090 | fs_info->sectorsize_bits; | |
2091 | int ret; | |
2092 | ||
2093 | /* The rbio should not have its csum buffer initialized. */ | |
2094 | ASSERT(!rbio->csum_buf && !rbio->csum_bitmap); | |
2095 | ||
2096 | /* | |
2097 | * Skip the csum search if: | |
2098 | * | |
2099 | * - The rbio doesn't belong to data block groups | |
2100 | * Then we are doing IO for tree blocks, no need to search csums. | |
2101 | * | |
2102 | * - The rbio belongs to mixed block groups | |
2103 | * This is to avoid deadlock, as we're already holding the full | |
2104 | * stripe lock, if we trigger a metadata read, and it needs to do | |
2105 | * raid56 recovery, we will deadlock. | |
2106 | */ | |
2107 | if (!(rbio->bioc->map_type & BTRFS_BLOCK_GROUP_DATA) || | |
2108 | rbio->bioc->map_type & BTRFS_BLOCK_GROUP_METADATA) | |
2109 | return; | |
2110 | ||
2111 | rbio->csum_buf = kzalloc(rbio->nr_data * rbio->stripe_nsectors * | |
2112 | fs_info->csum_size, GFP_NOFS); | |
2113 | rbio->csum_bitmap = bitmap_zalloc(rbio->nr_data * rbio->stripe_nsectors, | |
2114 | GFP_NOFS); | |
2115 | if (!rbio->csum_buf || !rbio->csum_bitmap) { | |
2116 | ret = -ENOMEM; | |
2117 | goto error; | |
2118 | } | |
2119 | ||
2120 | ret = btrfs_lookup_csums_bitmap(csum_root, start, start + len - 1, | |
2121 | rbio->csum_buf, rbio->csum_bitmap); | |
2122 | if (ret < 0) | |
2123 | goto error; | |
2124 | if (bitmap_empty(rbio->csum_bitmap, len >> fs_info->sectorsize_bits)) | |
2125 | goto no_csum; | |
2126 | return; | |
2127 | ||
2128 | error: | |
2129 | /* | |
2130 | * We failed to allocate memory or grab the csum, but it's not fatal, | |
2131 | * we can still continue. But better to warn users that RMW is no | |
2132 | * longer safe for this particular sub-stripe write. | |
2133 | */ | |
2134 | btrfs_warn_rl(fs_info, | |
2135 | "sub-stripe write for full stripe %llu is not safe, failed to get csum: %d", | |
2136 | rbio->bioc->raid_map[0], ret); | |
2137 | no_csum: | |
2138 | kfree(rbio->csum_buf); | |
2139 | bitmap_free(rbio->csum_bitmap); | |
2140 | rbio->csum_buf = NULL; | |
2141 | rbio->csum_bitmap = NULL; | |
2142 | } | |
2143 | ||
7a315072 | 2144 | static int rmw_read_wait_recover(struct btrfs_raid_bio *rbio) |
5eb30ee2 | 2145 | { |
02efa3a6 CH |
2146 | struct bio_list bio_list = BIO_EMPTY_LIST; |
2147 | int total_sector_nr; | |
2148 | int ret = 0; | |
5eb30ee2 | 2149 | |
c5a41562 QW |
2150 | /* |
2151 | * Fill the data csums we need for data verification. We need to fill | |
2152 | * the csum_bitmap/csum_buf first, as our endio function will try to | |
2153 | * verify the data sectors. | |
2154 | */ | |
2155 | fill_data_csums(rbio); | |
2156 | ||
02efa3a6 CH |
2157 | /* |
2158 | * Build a list of bios to read all sectors (including data and P/Q). | |
2159 | * | |
2160 | * This behavior is to compensate the later csum verification and recovery. | |
2161 | */ | |
2162 | for (total_sector_nr = 0; total_sector_nr < rbio->nr_sectors; | |
2163 | total_sector_nr++) { | |
2164 | struct sector_ptr *sector; | |
2165 | int stripe = total_sector_nr / rbio->stripe_nsectors; | |
2166 | int sectornr = total_sector_nr % rbio->stripe_nsectors; | |
5eb30ee2 | 2167 | |
02efa3a6 CH |
2168 | sector = rbio_stripe_sector(rbio, stripe, sectornr); |
2169 | ret = rbio_add_io_sector(rbio, &bio_list, sector, | |
2170 | stripe, sectornr, REQ_OP_READ); | |
2171 | if (ret) { | |
2172 | bio_list_put(&bio_list); | |
2173 | return ret; | |
2174 | } | |
2175 | } | |
7a315072 QW |
2176 | |
2177 | /* | |
2178 | * We may or may not have any corrupted sectors (including missing dev | |
2179 | * and csum mismatch), just let recover_sectors() to handle them all. | |
2180 | */ | |
02efa3a6 CH |
2181 | submit_read_wait_bio_list(rbio, &bio_list); |
2182 | return recover_sectors(rbio); | |
5eb30ee2 QW |
2183 | } |
2184 | ||
2185 | static void raid_wait_write_end_io(struct bio *bio) | |
2186 | { | |
2187 | struct btrfs_raid_bio *rbio = bio->bi_private; | |
2188 | blk_status_t err = bio->bi_status; | |
2189 | ||
ad3daf1c | 2190 | if (err) |
2942a50d | 2191 | rbio_update_error_bitmap(rbio, bio); |
5eb30ee2 QW |
2192 | bio_put(bio); |
2193 | if (atomic_dec_and_test(&rbio->stripes_pending)) | |
2194 | wake_up(&rbio->io_wait); | |
2195 | } | |
2196 | ||
2197 | static void submit_write_bios(struct btrfs_raid_bio *rbio, | |
2198 | struct bio_list *bio_list) | |
2199 | { | |
2200 | struct bio *bio; | |
2201 | ||
2202 | atomic_set(&rbio->stripes_pending, bio_list_size(bio_list)); | |
2203 | while ((bio = bio_list_pop(bio_list))) { | |
2204 | bio->bi_end_io = raid_wait_write_end_io; | |
2205 | ||
2206 | if (trace_raid56_write_stripe_enabled()) { | |
2207 | struct raid56_bio_trace_info trace_info = { 0 }; | |
2208 | ||
2209 | bio_get_trace_info(rbio, bio, &trace_info); | |
2210 | trace_raid56_write_stripe(rbio, bio, &trace_info); | |
2211 | } | |
2212 | submit_bio(bio); | |
2213 | } | |
2214 | } | |
2215 | ||
7a315072 QW |
2216 | /* |
2217 | * To determine if we need to read any sector from the disk. | |
2218 | * Should only be utilized in RMW path, to skip cached rbio. | |
2219 | */ | |
2220 | static bool need_read_stripe_sectors(struct btrfs_raid_bio *rbio) | |
2221 | { | |
2222 | int i; | |
2223 | ||
2224 | for (i = 0; i < rbio->nr_data * rbio->stripe_nsectors; i++) { | |
2225 | struct sector_ptr *sector = &rbio->stripe_sectors[i]; | |
2226 | ||
2227 | /* | |
2228 | * We have a sector which doesn't have page nor uptodate, | |
2229 | * thus this rbio can not be cached one, as cached one must | |
2230 | * have all its data sectors present and uptodate. | |
2231 | */ | |
2232 | if (!sector->page || !sector->uptodate) | |
2233 | return true; | |
2234 | } | |
2235 | return false; | |
2236 | } | |
2237 | ||
93723095 | 2238 | static int rmw_rbio(struct btrfs_raid_bio *rbio) |
5eb30ee2 QW |
2239 | { |
2240 | struct bio_list bio_list; | |
2241 | int sectornr; | |
2242 | int ret = 0; | |
2243 | ||
2244 | /* | |
2245 | * Allocate the pages for parity first, as P/Q pages will always be | |
2246 | * needed for both full-stripe and sub-stripe writes. | |
2247 | */ | |
2248 | ret = alloc_rbio_parity_pages(rbio); | |
2249 | if (ret < 0) | |
2250 | return ret; | |
2251 | ||
7a315072 QW |
2252 | /* |
2253 | * Either full stripe write, or we have every data sector already | |
2254 | * cached, can go to write path immediately. | |
2255 | */ | |
4d762701 CH |
2256 | if (!rbio_is_full(rbio) && need_read_stripe_sectors(rbio)) { |
2257 | /* | |
2258 | * Now we're doing sub-stripe write, also need all data stripes | |
2259 | * to do the full RMW. | |
2260 | */ | |
2261 | ret = alloc_rbio_data_pages(rbio); | |
2262 | if (ret < 0) | |
2263 | return ret; | |
5eb30ee2 | 2264 | |
4d762701 | 2265 | index_rbio_pages(rbio); |
5eb30ee2 | 2266 | |
4d762701 CH |
2267 | ret = rmw_read_wait_recover(rbio); |
2268 | if (ret < 0) | |
2269 | return ret; | |
2270 | } | |
5eb30ee2 | 2271 | |
5eb30ee2 QW |
2272 | /* |
2273 | * At this stage we're not allowed to add any new bios to the | |
2274 | * bio list any more, anyone else that wants to change this stripe | |
2275 | * needs to do their own rmw. | |
2276 | */ | |
2277 | spin_lock_irq(&rbio->bio_list_lock); | |
2278 | set_bit(RBIO_RMW_LOCKED_BIT, &rbio->flags); | |
2279 | spin_unlock_irq(&rbio->bio_list_lock); | |
2280 | ||
2942a50d | 2281 | bitmap_clear(rbio->error_bitmap, 0, rbio->nr_sectors); |
5eb30ee2 QW |
2282 | |
2283 | index_rbio_pages(rbio); | |
2284 | ||
2285 | /* | |
2286 | * We don't cache full rbios because we're assuming | |
2287 | * the higher layers are unlikely to use this area of | |
2288 | * the disk again soon. If they do use it again, | |
2289 | * hopefully they will send another full bio. | |
2290 | */ | |
2291 | if (!rbio_is_full(rbio)) | |
2292 | cache_rbio_pages(rbio); | |
2293 | else | |
2294 | clear_bit(RBIO_CACHE_READY_BIT, &rbio->flags); | |
2295 | ||
2296 | for (sectornr = 0; sectornr < rbio->stripe_nsectors; sectornr++) | |
2297 | generate_pq_vertical(rbio, sectornr); | |
2298 | ||
2299 | bio_list_init(&bio_list); | |
2300 | ret = rmw_assemble_write_bios(rbio, &bio_list); | |
2301 | if (ret < 0) | |
2302 | return ret; | |
2303 | ||
2304 | /* We should have at least one bio assembled. */ | |
2305 | ASSERT(bio_list_size(&bio_list)); | |
2306 | submit_write_bios(rbio, &bio_list); | |
2307 | wait_event(rbio->io_wait, atomic_read(&rbio->stripes_pending) == 0); | |
2308 | ||
ad3daf1c QW |
2309 | /* We may have more errors than our tolerance during the read. */ |
2310 | for (sectornr = 0; sectornr < rbio->stripe_nsectors; sectornr++) { | |
2311 | int found_errors; | |
2312 | ||
2313 | found_errors = get_rbio_veritical_errors(rbio, sectornr, NULL, NULL); | |
2314 | if (found_errors > rbio->bioc->max_errors) { | |
2315 | ret = -EIO; | |
2316 | break; | |
2317 | } | |
2318 | } | |
5eb30ee2 QW |
2319 | return ret; |
2320 | } | |
2321 | ||
93723095 QW |
2322 | static void rmw_rbio_work(struct work_struct *work) |
2323 | { | |
2324 | struct btrfs_raid_bio *rbio; | |
2325 | int ret; | |
2326 | ||
2327 | rbio = container_of(work, struct btrfs_raid_bio, work); | |
2328 | ||
2329 | ret = lock_stripe_add(rbio); | |
2330 | if (ret == 0) { | |
2331 | ret = rmw_rbio(rbio); | |
2332 | rbio_orig_end_io(rbio, errno_to_blk_status(ret)); | |
2333 | } | |
2334 | } | |
2335 | ||
2336 | static void rmw_rbio_work_locked(struct work_struct *work) | |
53b381b3 DW |
2337 | { |
2338 | struct btrfs_raid_bio *rbio; | |
93723095 | 2339 | int ret; |
53b381b3 DW |
2340 | |
2341 | rbio = container_of(work, struct btrfs_raid_bio, work); | |
93723095 QW |
2342 | |
2343 | ret = rmw_rbio(rbio); | |
2344 | rbio_orig_end_io(rbio, errno_to_blk_status(ret)); | |
53b381b3 DW |
2345 | } |
2346 | ||
5a6ac9ea MX |
2347 | /* |
2348 | * The following code is used to scrub/replace the parity stripe | |
2349 | * | |
4c664611 | 2350 | * Caller must have already increased bio_counter for getting @bioc. |
ae6529c3 | 2351 | * |
5a6ac9ea MX |
2352 | * Note: We need make sure all the pages that add into the scrub/replace |
2353 | * raid bio are correct and not be changed during the scrub/replace. That | |
2354 | * is those pages just hold metadata or file data with checksum. | |
2355 | */ | |
2356 | ||
6a258d72 QW |
2357 | struct btrfs_raid_bio *raid56_parity_alloc_scrub_rbio(struct bio *bio, |
2358 | struct btrfs_io_context *bioc, | |
ff18a4af | 2359 | struct btrfs_device *scrub_dev, |
6a258d72 | 2360 | unsigned long *dbitmap, int stripe_nsectors) |
5a6ac9ea | 2361 | { |
6a258d72 | 2362 | struct btrfs_fs_info *fs_info = bioc->fs_info; |
5a6ac9ea MX |
2363 | struct btrfs_raid_bio *rbio; |
2364 | int i; | |
2365 | ||
ff18a4af | 2366 | rbio = alloc_rbio(fs_info, bioc); |
5a6ac9ea MX |
2367 | if (IS_ERR(rbio)) |
2368 | return NULL; | |
2369 | bio_list_add(&rbio->bio_list, bio); | |
2370 | /* | |
2371 | * This is a special bio which is used to hold the completion handler | |
2372 | * and make the scrub rbio is similar to the other types | |
2373 | */ | |
2374 | ASSERT(!bio->bi_iter.bi_size); | |
2375 | rbio->operation = BTRFS_RBIO_PARITY_SCRUB; | |
2376 | ||
9cd3a7eb | 2377 | /* |
4c664611 | 2378 | * After mapping bioc with BTRFS_MAP_WRITE, parities have been sorted |
9cd3a7eb LB |
2379 | * to the end position, so this search can start from the first parity |
2380 | * stripe. | |
2381 | */ | |
2382 | for (i = rbio->nr_data; i < rbio->real_stripes; i++) { | |
4c664611 | 2383 | if (bioc->stripes[i].dev == scrub_dev) { |
5a6ac9ea MX |
2384 | rbio->scrubp = i; |
2385 | break; | |
2386 | } | |
2387 | } | |
9cd3a7eb | 2388 | ASSERT(i < rbio->real_stripes); |
5a6ac9ea | 2389 | |
c67c68eb | 2390 | bitmap_copy(&rbio->dbitmap, dbitmap, stripe_nsectors); |
5a6ac9ea MX |
2391 | return rbio; |
2392 | } | |
2393 | ||
b4ee1782 OS |
2394 | /* Used for both parity scrub and missing. */ |
2395 | void raid56_add_scrub_pages(struct btrfs_raid_bio *rbio, struct page *page, | |
6346f6bf | 2396 | unsigned int pgoff, u64 logical) |
5a6ac9ea | 2397 | { |
6346f6bf | 2398 | const u32 sectorsize = rbio->bioc->fs_info->sectorsize; |
5a6ac9ea MX |
2399 | int stripe_offset; |
2400 | int index; | |
2401 | ||
4c664611 | 2402 | ASSERT(logical >= rbio->bioc->raid_map[0]); |
6346f6bf | 2403 | ASSERT(logical + sectorsize <= rbio->bioc->raid_map[0] + |
ff18a4af | 2404 | BTRFS_STRIPE_LEN * rbio->nr_data); |
4c664611 | 2405 | stripe_offset = (int)(logical - rbio->bioc->raid_map[0]); |
6346f6bf QW |
2406 | index = stripe_offset / sectorsize; |
2407 | rbio->bio_sectors[index].page = page; | |
2408 | rbio->bio_sectors[index].pgoff = pgoff; | |
5a6ac9ea MX |
2409 | } |
2410 | ||
2411 | /* | |
2412 | * We just scrub the parity that we have correct data on the same horizontal, | |
2413 | * so we needn't allocate all pages for all the stripes. | |
2414 | */ | |
2415 | static int alloc_rbio_essential_pages(struct btrfs_raid_bio *rbio) | |
2416 | { | |
3907ce29 | 2417 | const u32 sectorsize = rbio->bioc->fs_info->sectorsize; |
aee35e4b | 2418 | int total_sector_nr; |
5a6ac9ea | 2419 | |
aee35e4b QW |
2420 | for (total_sector_nr = 0; total_sector_nr < rbio->nr_sectors; |
2421 | total_sector_nr++) { | |
2422 | struct page *page; | |
2423 | int sectornr = total_sector_nr % rbio->stripe_nsectors; | |
2424 | int index = (total_sector_nr * sectorsize) >> PAGE_SHIFT; | |
5a6ac9ea | 2425 | |
aee35e4b QW |
2426 | if (!test_bit(sectornr, &rbio->dbitmap)) |
2427 | continue; | |
2428 | if (rbio->stripe_pages[index]) | |
2429 | continue; | |
2430 | page = alloc_page(GFP_NOFS); | |
2431 | if (!page) | |
2432 | return -ENOMEM; | |
2433 | rbio->stripe_pages[index] = page; | |
5a6ac9ea | 2434 | } |
eb357060 | 2435 | index_stripe_sectors(rbio); |
5a6ac9ea MX |
2436 | return 0; |
2437 | } | |
2438 | ||
6bfd0133 | 2439 | static int finish_parity_scrub(struct btrfs_raid_bio *rbio, int need_check) |
5a6ac9ea | 2440 | { |
4c664611 | 2441 | struct btrfs_io_context *bioc = rbio->bioc; |
46900662 | 2442 | const u32 sectorsize = bioc->fs_info->sectorsize; |
1389053e | 2443 | void **pointers = rbio->finish_pointers; |
c67c68eb | 2444 | unsigned long *pbitmap = &rbio->finish_pbitmap; |
5a6ac9ea MX |
2445 | int nr_data = rbio->nr_data; |
2446 | int stripe; | |
3e77605d | 2447 | int sectornr; |
c17af965 | 2448 | bool has_qstripe; |
46900662 QW |
2449 | struct sector_ptr p_sector = { 0 }; |
2450 | struct sector_ptr q_sector = { 0 }; | |
5a6ac9ea | 2451 | struct bio_list bio_list; |
76035976 | 2452 | int is_replace = 0; |
5a6ac9ea MX |
2453 | int ret; |
2454 | ||
2455 | bio_list_init(&bio_list); | |
2456 | ||
c17af965 DS |
2457 | if (rbio->real_stripes - rbio->nr_data == 1) |
2458 | has_qstripe = false; | |
2459 | else if (rbio->real_stripes - rbio->nr_data == 2) | |
2460 | has_qstripe = true; | |
2461 | else | |
5a6ac9ea | 2462 | BUG(); |
5a6ac9ea | 2463 | |
4c664611 | 2464 | if (bioc->num_tgtdevs && bioc->tgtdev_map[rbio->scrubp]) { |
76035976 | 2465 | is_replace = 1; |
c67c68eb | 2466 | bitmap_copy(pbitmap, &rbio->dbitmap, rbio->stripe_nsectors); |
76035976 MX |
2467 | } |
2468 | ||
5a6ac9ea MX |
2469 | /* |
2470 | * Because the higher layers(scrubber) are unlikely to | |
2471 | * use this area of the disk again soon, so don't cache | |
2472 | * it. | |
2473 | */ | |
2474 | clear_bit(RBIO_CACHE_READY_BIT, &rbio->flags); | |
2475 | ||
2476 | if (!need_check) | |
2477 | goto writeback; | |
2478 | ||
46900662 QW |
2479 | p_sector.page = alloc_page(GFP_NOFS); |
2480 | if (!p_sector.page) | |
6bfd0133 | 2481 | return -ENOMEM; |
46900662 QW |
2482 | p_sector.pgoff = 0; |
2483 | p_sector.uptodate = 1; | |
5a6ac9ea | 2484 | |
c17af965 | 2485 | if (has_qstripe) { |
d70cef0d | 2486 | /* RAID6, allocate and map temp space for the Q stripe */ |
46900662 QW |
2487 | q_sector.page = alloc_page(GFP_NOFS); |
2488 | if (!q_sector.page) { | |
2489 | __free_page(p_sector.page); | |
2490 | p_sector.page = NULL; | |
6bfd0133 | 2491 | return -ENOMEM; |
5a6ac9ea | 2492 | } |
46900662 QW |
2493 | q_sector.pgoff = 0; |
2494 | q_sector.uptodate = 1; | |
2495 | pointers[rbio->real_stripes - 1] = kmap_local_page(q_sector.page); | |
5a6ac9ea MX |
2496 | } |
2497 | ||
2942a50d | 2498 | bitmap_clear(rbio->error_bitmap, 0, rbio->nr_sectors); |
5a6ac9ea | 2499 | |
d70cef0d | 2500 | /* Map the parity stripe just once */ |
46900662 | 2501 | pointers[nr_data] = kmap_local_page(p_sector.page); |
d70cef0d | 2502 | |
c67c68eb | 2503 | for_each_set_bit(sectornr, &rbio->dbitmap, rbio->stripe_nsectors) { |
46900662 | 2504 | struct sector_ptr *sector; |
5a6ac9ea | 2505 | void *parity; |
46900662 | 2506 | |
5a6ac9ea MX |
2507 | /* first collect one page from each data stripe */ |
2508 | for (stripe = 0; stripe < nr_data; stripe++) { | |
46900662 QW |
2509 | sector = sector_in_rbio(rbio, stripe, sectornr, 0); |
2510 | pointers[stripe] = kmap_local_page(sector->page) + | |
2511 | sector->pgoff; | |
5a6ac9ea MX |
2512 | } |
2513 | ||
c17af965 | 2514 | if (has_qstripe) { |
d70cef0d | 2515 | /* RAID6, call the library function to fill in our P/Q */ |
46900662 | 2516 | raid6_call.gen_syndrome(rbio->real_stripes, sectorsize, |
5a6ac9ea MX |
2517 | pointers); |
2518 | } else { | |
2519 | /* raid5 */ | |
46900662 QW |
2520 | memcpy(pointers[nr_data], pointers[0], sectorsize); |
2521 | run_xor(pointers + 1, nr_data - 1, sectorsize); | |
5a6ac9ea MX |
2522 | } |
2523 | ||
01327610 | 2524 | /* Check scrubbing parity and repair it */ |
46900662 QW |
2525 | sector = rbio_stripe_sector(rbio, rbio->scrubp, sectornr); |
2526 | parity = kmap_local_page(sector->page) + sector->pgoff; | |
2527 | if (memcmp(parity, pointers[rbio->scrubp], sectorsize) != 0) | |
2528 | memcpy(parity, pointers[rbio->scrubp], sectorsize); | |
5a6ac9ea MX |
2529 | else |
2530 | /* Parity is right, needn't writeback */ | |
c67c68eb | 2531 | bitmap_clear(&rbio->dbitmap, sectornr, 1); |
58c1a35c | 2532 | kunmap_local(parity); |
5a6ac9ea | 2533 | |
94a0b58d IW |
2534 | for (stripe = nr_data - 1; stripe >= 0; stripe--) |
2535 | kunmap_local(pointers[stripe]); | |
5a6ac9ea MX |
2536 | } |
2537 | ||
94a0b58d | 2538 | kunmap_local(pointers[nr_data]); |
46900662 QW |
2539 | __free_page(p_sector.page); |
2540 | p_sector.page = NULL; | |
2541 | if (q_sector.page) { | |
94a0b58d | 2542 | kunmap_local(pointers[rbio->real_stripes - 1]); |
46900662 QW |
2543 | __free_page(q_sector.page); |
2544 | q_sector.page = NULL; | |
d70cef0d | 2545 | } |
5a6ac9ea MX |
2546 | |
2547 | writeback: | |
2548 | /* | |
2549 | * time to start writing. Make bios for everything from the | |
2550 | * higher layers (the bio_list in our rbio) and our p/q. Ignore | |
2551 | * everything else. | |
2552 | */ | |
c67c68eb | 2553 | for_each_set_bit(sectornr, &rbio->dbitmap, rbio->stripe_nsectors) { |
3e77605d | 2554 | struct sector_ptr *sector; |
5a6ac9ea | 2555 | |
3e77605d QW |
2556 | sector = rbio_stripe_sector(rbio, rbio->scrubp, sectornr); |
2557 | ret = rbio_add_io_sector(rbio, &bio_list, sector, rbio->scrubp, | |
ff18a4af | 2558 | sectornr, REQ_OP_WRITE); |
5a6ac9ea MX |
2559 | if (ret) |
2560 | goto cleanup; | |
2561 | } | |
2562 | ||
76035976 MX |
2563 | if (!is_replace) |
2564 | goto submit_write; | |
2565 | ||
3e77605d QW |
2566 | for_each_set_bit(sectornr, pbitmap, rbio->stripe_nsectors) { |
2567 | struct sector_ptr *sector; | |
76035976 | 2568 | |
3e77605d QW |
2569 | sector = rbio_stripe_sector(rbio, rbio->scrubp, sectornr); |
2570 | ret = rbio_add_io_sector(rbio, &bio_list, sector, | |
4c664611 | 2571 | bioc->tgtdev_map[rbio->scrubp], |
ff18a4af | 2572 | sectornr, REQ_OP_WRITE); |
76035976 MX |
2573 | if (ret) |
2574 | goto cleanup; | |
2575 | } | |
2576 | ||
2577 | submit_write: | |
6bfd0133 QW |
2578 | submit_write_bios(rbio, &bio_list); |
2579 | return 0; | |
5a6ac9ea MX |
2580 | |
2581 | cleanup: | |
801fcfc5 | 2582 | bio_list_put(&bio_list); |
6bfd0133 | 2583 | return ret; |
5a6ac9ea MX |
2584 | } |
2585 | ||
2586 | static inline int is_data_stripe(struct btrfs_raid_bio *rbio, int stripe) | |
2587 | { | |
2588 | if (stripe >= 0 && stripe < rbio->nr_data) | |
2589 | return 1; | |
2590 | return 0; | |
2591 | } | |
2592 | ||
6bfd0133 | 2593 | static int recover_scrub_rbio(struct btrfs_raid_bio *rbio) |
5a6ac9ea | 2594 | { |
75b47033 QW |
2595 | void **pointers = NULL; |
2596 | void **unmap_array = NULL; | |
2597 | int sector_nr; | |
e7fc357e | 2598 | int ret = 0; |
5a6ac9ea | 2599 | |
75b47033 QW |
2600 | /* |
2601 | * @pointers array stores the pointer for each sector. | |
2602 | * | |
2603 | * @unmap_array stores copy of pointers that does not get reordered | |
2604 | * during reconstruction so that kunmap_local works. | |
2605 | */ | |
2606 | pointers = kcalloc(rbio->real_stripes, sizeof(void *), GFP_NOFS); | |
2607 | unmap_array = kcalloc(rbio->real_stripes, sizeof(void *), GFP_NOFS); | |
2608 | if (!pointers || !unmap_array) { | |
2609 | ret = -ENOMEM; | |
2610 | goto out; | |
2611 | } | |
5a6ac9ea | 2612 | |
75b47033 QW |
2613 | for (sector_nr = 0; sector_nr < rbio->stripe_nsectors; sector_nr++) { |
2614 | int dfail = 0, failp = -1; | |
2615 | int faila; | |
2616 | int failb; | |
2617 | int found_errors; | |
5a6ac9ea | 2618 | |
75b47033 QW |
2619 | found_errors = get_rbio_veritical_errors(rbio, sector_nr, |
2620 | &faila, &failb); | |
2621 | if (found_errors > rbio->bioc->max_errors) { | |
2622 | ret = -EIO; | |
2623 | goto out; | |
2624 | } | |
2625 | if (found_errors == 0) | |
2626 | continue; | |
5a6ac9ea | 2627 | |
75b47033 QW |
2628 | /* We should have at least one error here. */ |
2629 | ASSERT(faila >= 0 || failb >= 0); | |
5a6ac9ea | 2630 | |
75b47033 QW |
2631 | if (is_data_stripe(rbio, faila)) |
2632 | dfail++; | |
2633 | else if (is_parity_stripe(faila)) | |
2634 | failp = faila; | |
5a6ac9ea | 2635 | |
75b47033 QW |
2636 | if (is_data_stripe(rbio, failb)) |
2637 | dfail++; | |
2638 | else if (is_parity_stripe(failb)) | |
2639 | failp = failb; | |
2640 | /* | |
2641 | * Because we can not use a scrubbing parity to repair the | |
2642 | * data, so the capability of the repair is declined. (In the | |
2643 | * case of RAID5, we can not repair anything.) | |
2644 | */ | |
2645 | if (dfail > rbio->bioc->max_errors - 1) { | |
2646 | ret = -EIO; | |
2647 | goto out; | |
2648 | } | |
2649 | /* | |
2650 | * If all data is good, only parity is correctly, just repair | |
2651 | * the parity, no need to recover data stripes. | |
2652 | */ | |
2653 | if (dfail == 0) | |
2654 | continue; | |
6bfd0133 | 2655 | |
75b47033 QW |
2656 | /* |
2657 | * Here means we got one corrupted data stripe and one | |
2658 | * corrupted parity on RAID6, if the corrupted parity is | |
2659 | * scrubbing parity, luckily, use the other one to repair the | |
2660 | * data, or we can not repair the data stripe. | |
2661 | */ | |
2662 | if (failp != rbio->scrubp) { | |
2663 | ret = -EIO; | |
2664 | goto out; | |
2665 | } | |
2666 | ||
2667 | ret = recover_vertical(rbio, sector_nr, pointers, unmap_array); | |
2668 | if (ret < 0) | |
2669 | goto out; | |
2670 | } | |
2671 | out: | |
2672 | kfree(pointers); | |
2673 | kfree(unmap_array); | |
6bfd0133 | 2674 | return ret; |
5a6ac9ea MX |
2675 | } |
2676 | ||
52f0c198 | 2677 | static int scrub_assemble_read_bios(struct btrfs_raid_bio *rbio) |
5a6ac9ea | 2678 | { |
52f0c198 | 2679 | struct bio_list bio_list = BIO_EMPTY_LIST; |
cb3450b7 QW |
2680 | int total_sector_nr; |
2681 | int ret = 0; | |
5a6ac9ea | 2682 | |
1c10702e QW |
2683 | /* Build a list of bios to read all the missing parts. */ |
2684 | for (total_sector_nr = 0; total_sector_nr < rbio->nr_sectors; | |
2685 | total_sector_nr++) { | |
2686 | int sectornr = total_sector_nr % rbio->stripe_nsectors; | |
2687 | int stripe = total_sector_nr / rbio->stripe_nsectors; | |
2688 | struct sector_ptr *sector; | |
5a6ac9ea | 2689 | |
1c10702e QW |
2690 | /* No data in the vertical stripe, no need to read. */ |
2691 | if (!test_bit(sectornr, &rbio->dbitmap)) | |
2692 | continue; | |
5a6ac9ea | 2693 | |
1c10702e QW |
2694 | /* |
2695 | * We want to find all the sectors missing from the rbio and | |
2696 | * read them from the disk. If sector_in_rbio() finds a sector | |
2697 | * in the bio list we don't need to read it off the stripe. | |
2698 | */ | |
2699 | sector = sector_in_rbio(rbio, stripe, sectornr, 1); | |
2700 | if (sector) | |
2701 | continue; | |
2702 | ||
2703 | sector = rbio_stripe_sector(rbio, stripe, sectornr); | |
2704 | /* | |
2705 | * The bio cache may have handed us an uptodate sector. If so, | |
2706 | * use it. | |
2707 | */ | |
2708 | if (sector->uptodate) | |
2709 | continue; | |
2710 | ||
52f0c198 | 2711 | ret = rbio_add_io_sector(rbio, &bio_list, sector, stripe, |
ff18a4af | 2712 | sectornr, REQ_OP_READ); |
52f0c198 CH |
2713 | if (ret) { |
2714 | bio_list_put(&bio_list); | |
2715 | return ret; | |
2716 | } | |
5a6ac9ea | 2717 | } |
52f0c198 CH |
2718 | |
2719 | submit_read_wait_bio_list(rbio, &bio_list); | |
cb3450b7 | 2720 | return 0; |
cb3450b7 QW |
2721 | } |
2722 | ||
6bfd0133 | 2723 | static int scrub_rbio(struct btrfs_raid_bio *rbio) |
cb3450b7 | 2724 | { |
6bfd0133 | 2725 | bool need_check = false; |
ad3daf1c | 2726 | int sector_nr; |
cb3450b7 | 2727 | int ret; |
cb3450b7 | 2728 | |
cb3450b7 QW |
2729 | ret = alloc_rbio_essential_pages(rbio); |
2730 | if (ret) | |
52f0c198 | 2731 | return ret; |
cb3450b7 | 2732 | |
2942a50d QW |
2733 | bitmap_clear(rbio->error_bitmap, 0, rbio->nr_sectors); |
2734 | ||
52f0c198 | 2735 | ret = scrub_assemble_read_bios(rbio); |
cb3450b7 | 2736 | if (ret < 0) |
52f0c198 | 2737 | return ret; |
5a6ac9ea | 2738 | |
75b47033 | 2739 | /* We may have some failures, recover the failed sectors first. */ |
6bfd0133 QW |
2740 | ret = recover_scrub_rbio(rbio); |
2741 | if (ret < 0) | |
52f0c198 | 2742 | return ret; |
5a6ac9ea | 2743 | |
6bfd0133 QW |
2744 | /* |
2745 | * We have every sector properly prepared. Can finish the scrub | |
2746 | * and writeback the good content. | |
2747 | */ | |
2748 | ret = finish_parity_scrub(rbio, need_check); | |
2749 | wait_event(rbio->io_wait, atomic_read(&rbio->stripes_pending) == 0); | |
ad3daf1c QW |
2750 | for (sector_nr = 0; sector_nr < rbio->stripe_nsectors; sector_nr++) { |
2751 | int found_errors; | |
2752 | ||
2753 | found_errors = get_rbio_veritical_errors(rbio, sector_nr, NULL, NULL); | |
2754 | if (found_errors > rbio->bioc->max_errors) { | |
2755 | ret = -EIO; | |
2756 | break; | |
2757 | } | |
2758 | } | |
6bfd0133 | 2759 | return ret; |
5a6ac9ea MX |
2760 | } |
2761 | ||
6bfd0133 | 2762 | static void scrub_rbio_work_locked(struct work_struct *work) |
5a6ac9ea MX |
2763 | { |
2764 | struct btrfs_raid_bio *rbio; | |
6bfd0133 | 2765 | int ret; |
5a6ac9ea MX |
2766 | |
2767 | rbio = container_of(work, struct btrfs_raid_bio, work); | |
6bfd0133 QW |
2768 | ret = scrub_rbio(rbio); |
2769 | rbio_orig_end_io(rbio, errno_to_blk_status(ret)); | |
5a6ac9ea MX |
2770 | } |
2771 | ||
5a6ac9ea MX |
2772 | void raid56_parity_submit_scrub_rbio(struct btrfs_raid_bio *rbio) |
2773 | { | |
2774 | if (!lock_stripe_add(rbio)) | |
6bfd0133 | 2775 | start_async_work(rbio, scrub_rbio_work_locked); |
5a6ac9ea | 2776 | } |
b4ee1782 OS |
2777 | |
2778 | /* The following code is used for dev replace of a missing RAID 5/6 device. */ | |
2779 | ||
2780 | struct btrfs_raid_bio * | |
ff18a4af | 2781 | raid56_alloc_missing_rbio(struct bio *bio, struct btrfs_io_context *bioc) |
b4ee1782 | 2782 | { |
6a258d72 | 2783 | struct btrfs_fs_info *fs_info = bioc->fs_info; |
b4ee1782 OS |
2784 | struct btrfs_raid_bio *rbio; |
2785 | ||
ff18a4af | 2786 | rbio = alloc_rbio(fs_info, bioc); |
b4ee1782 OS |
2787 | if (IS_ERR(rbio)) |
2788 | return NULL; | |
2789 | ||
2790 | rbio->operation = BTRFS_RBIO_REBUILD_MISSING; | |
2791 | bio_list_add(&rbio->bio_list, bio); | |
2792 | /* | |
2793 | * This is a special bio which is used to hold the completion handler | |
2794 | * and make the scrub rbio is similar to the other types | |
2795 | */ | |
2796 | ASSERT(!bio->bi_iter.bi_size); | |
2797 | ||
2942a50d | 2798 | set_rbio_range_error(rbio, bio); |
b4ee1782 OS |
2799 | |
2800 | return rbio; | |
2801 | } | |
2802 | ||
b4ee1782 OS |
2803 | void raid56_submit_missing_rbio(struct btrfs_raid_bio *rbio) |
2804 | { | |
d817ce35 | 2805 | start_async_work(rbio, recover_rbio_work); |
b4ee1782 | 2806 | } |