1 | // SPDX-License-Identifier: GPL-2.0 |
2 | /* |
3 | * Copyright (C) 2012 Fusion-io All rights reserved. |
4 | * Copyright (C) 2012 Intel Corp. All rights reserved. |
5 | */ |
6 | |
7 | #include <linux/sched.h> |
8 | #include <linux/bio.h> |
9 | #include <linux/slab.h> |
10 | #include <linux/blkdev.h> |
11 | #include <linux/raid/pq.h> |
12 | #include <linux/hash.h> |
13 | #include <linux/list_sort.h> |
14 | #include <linux/raid/xor.h> |
15 | #include <linux/mm.h> |
16 | #include "messages.h" |
17 | #include "misc.h" |
18 | #include "ctree.h" |
19 | #include "disk-io.h" |
20 | #include "volumes.h" |
21 | #include "raid56.h" |
22 | #include "async-thread.h" |
23 | #include "file-item.h" |
24 | #include "btrfs_inode.h" |
25 | |
26 | /* set when additional merges to this rbio are not allowed */ |
27 | #define RBIO_RMW_LOCKED_BIT 1 |
28 | |
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 | |
40 | #define RBIO_CACHE_SIZE 1024 |
41 | |
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 | |
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; |
65 | unsigned int pgoff:24; |
66 | unsigned int uptodate:8; |
67 | }; |
68 | |
69 | static void rmw_rbio_work(struct work_struct *work); |
70 | static void rmw_rbio_work_locked(struct work_struct *work); |
71 | static void index_rbio_pages(struct btrfs_raid_bio *rbio); |
72 | static int alloc_rbio_pages(struct btrfs_raid_bio *rbio); |
73 | |
74 | static int finish_parity_scrub(struct btrfs_raid_bio *rbio); |
75 | static void scrub_rbio_work_locked(struct work_struct *work); |
76 | |
77 | static void free_raid_bio_pointers(struct btrfs_raid_bio *rbio) |
78 | { |
79 | bitmap_free(bitmap: rbio->error_bitmap); |
80 | kfree(objp: rbio->stripe_pages); |
81 | kfree(objp: rbio->bio_sectors); |
82 | kfree(objp: rbio->stripe_sectors); |
83 | kfree(objp: rbio->finish_pointers); |
84 | } |
85 | |
86 | static void free_raid_bio(struct btrfs_raid_bio *rbio) |
87 | { |
88 | int i; |
89 | |
90 | if (!refcount_dec_and_test(r: &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(bioc: rbio->bioc); |
105 | free_raid_bio_pointers(rbio); |
106 | kfree(objp: rbio); |
107 | } |
108 | |
109 | static void start_async_work(struct btrfs_raid_bio *rbio, work_func_t work_func) |
110 | { |
111 | INIT_WORK(&rbio->work, work_func); |
112 | queue_work(wq: rbio->bioc->fs_info->rmw_workers, work: &rbio->work); |
113 | } |
114 | |
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 | |
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 | */ |
138 | table = kvzalloc(struct_size(table, table, num_entries), GFP_KERNEL); |
139 | if (!table) |
140 | return -ENOMEM; |
141 | |
142 | spin_lock_init(&table->cache_lock); |
143 | INIT_LIST_HEAD(list: &table->stripe_cache); |
144 | |
145 | h = table->table; |
146 | |
147 | for (i = 0; i < num_entries; i++) { |
148 | cur = h + i; |
149 | INIT_LIST_HEAD(list: &cur->hash_list); |
150 | spin_lock_init(&cur->lock); |
151 | } |
152 | |
153 | x = cmpxchg(&info->stripe_hash_table, NULL, table); |
154 | kvfree(addr: x); |
155 | return 0; |
156 | } |
157 | |
158 | /* |
159 | * caching an rbio means to copy anything from the |
160 | * bio_sectors array into the stripe_pages array. We |
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; |
170 | int ret; |
171 | |
172 | ret = alloc_rbio_pages(rbio); |
173 | if (ret) |
174 | return; |
175 | |
176 | for (i = 0; i < rbio->nr_sectors; i++) { |
177 | /* Some range not covered by bio (partial write), skip it */ |
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); |
186 | continue; |
187 | } |
188 | |
189 | ASSERT(rbio->stripe_sectors[i].page); |
190 | memcpy_page(dst_page: rbio->stripe_sectors[i].page, |
191 | dst_off: rbio->stripe_sectors[i].pgoff, |
192 | src_page: rbio->bio_sectors[i].page, |
193 | src_off: rbio->bio_sectors[i].pgoff, |
194 | len: rbio->bioc->fs_info->sectorsize); |
195 | rbio->stripe_sectors[i].uptodate = 1; |
196 | } |
197 | set_bit(RBIO_CACHE_READY_BIT, addr: &rbio->flags); |
198 | } |
199 | |
200 | /* |
201 | * we hash on the first logical address of the stripe |
202 | */ |
203 | static int rbio_bucket(struct btrfs_raid_bio *rbio) |
204 | { |
205 | u64 num = rbio->bioc->full_stripe_logical; |
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(val: num >> 16, BTRFS_STRIPE_HASH_TABLE_BITS); |
216 | } |
217 | |
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 | |
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 | |
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 | |
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 | |
289 | /* |
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. |
295 | */ |
296 | static void steal_rbio(struct btrfs_raid_bio *src, struct btrfs_raid_bio *dest) |
297 | { |
298 | int i; |
299 | |
300 | if (!test_bit(RBIO_CACHE_READY_BIT, &src->flags)) |
301 | return; |
302 | |
303 | for (i = 0; i < dest->nr_pages; i++) { |
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(rbio: src, page_nr: i)) |
311 | continue; |
312 | |
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)); |
319 | steal_rbio_page(src, dest, page_nr: i); |
320 | } |
321 | index_stripe_sectors(rbio: dest); |
322 | index_stripe_sectors(rbio: src); |
323 | } |
324 | |
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(bl: &dest->bio_list, bl2: &victim->bio_list); |
336 | dest->bio_list_bytes += victim->bio_list_bytes; |
337 | /* Also inherit the bitmaps from @victim. */ |
338 | bitmap_or(dst: &dest->dbitmap, src1: &victim->dbitmap, src2: &dest->dbitmap, |
339 | nbits: dest->stripe_nsectors); |
340 | bio_list_init(bl: &victim->bio_list); |
341 | } |
342 | |
343 | /* |
344 | * used to prune items that are in the cache. The caller |
345 | * must hold the hash table lock. |
346 | */ |
347 | static void __remove_rbio_from_cache(struct btrfs_raid_bio *rbio) |
348 | { |
349 | int bucket = rbio_bucket(rbio); |
350 | struct btrfs_stripe_hash_table *table; |
351 | struct btrfs_stripe_hash *h; |
352 | int freeit = 0; |
353 | |
354 | /* |
355 | * check the bit again under the hash table lock. |
356 | */ |
357 | if (!test_bit(RBIO_CACHE_BIT, &rbio->flags)) |
358 | return; |
359 | |
360 | table = rbio->bioc->fs_info->stripe_hash_table; |
361 | h = table->table + bucket; |
362 | |
363 | /* hold the lock for the bucket because we may be |
364 | * removing it from the hash table |
365 | */ |
366 | spin_lock(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(lock: &rbio->bio_list_lock); |
373 | |
374 | if (test_and_clear_bit(RBIO_CACHE_BIT, addr: &rbio->flags)) { |
375 | list_del_init(entry: &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(bl: &rbio->bio_list)) { |
389 | if (!list_empty(head: &rbio->hash_list)) { |
390 | list_del_init(entry: &rbio->hash_list); |
391 | refcount_dec(r: &rbio->refs); |
392 | BUG_ON(!list_empty(&rbio->plug_list)); |
393 | } |
394 | } |
395 | } |
396 | |
397 | spin_unlock(lock: &rbio->bio_list_lock); |
398 | spin_unlock(lock: &h->lock); |
399 | |
400 | if (freeit) |
401 | free_raid_bio(rbio); |
402 | } |
403 | |
404 | /* |
405 | * prune a given rbio from the cache |
406 | */ |
407 | static void remove_rbio_from_cache(struct btrfs_raid_bio *rbio) |
408 | { |
409 | struct btrfs_stripe_hash_table *table; |
410 | |
411 | if (!test_bit(RBIO_CACHE_BIT, &rbio->flags)) |
412 | return; |
413 | |
414 | table = rbio->bioc->fs_info->stripe_hash_table; |
415 | |
416 | spin_lock(lock: &table->cache_lock); |
417 | __remove_rbio_from_cache(rbio); |
418 | spin_unlock(lock: &table->cache_lock); |
419 | } |
420 | |
421 | /* |
422 | * remove everything in the cache |
423 | */ |
424 | static void btrfs_clear_rbio_cache(struct btrfs_fs_info *info) |
425 | { |
426 | struct btrfs_stripe_hash_table *table; |
427 | struct btrfs_raid_bio *rbio; |
428 | |
429 | table = info->stripe_hash_table; |
430 | |
431 | spin_lock(lock: &table->cache_lock); |
432 | while (!list_empty(head: &table->stripe_cache)) { |
433 | rbio = list_entry(table->stripe_cache.next, |
434 | struct btrfs_raid_bio, |
435 | stripe_cache); |
436 | __remove_rbio_from_cache(rbio); |
437 | } |
438 | spin_unlock(lock: &table->cache_lock); |
439 | } |
440 | |
441 | /* |
442 | * remove all cached entries and free the hash table |
443 | * used by unmount |
444 | */ |
445 | void btrfs_free_stripe_hash_table(struct btrfs_fs_info *info) |
446 | { |
447 | if (!info->stripe_hash_table) |
448 | return; |
449 | btrfs_clear_rbio_cache(info); |
450 | kvfree(addr: info->stripe_hash_table); |
451 | info->stripe_hash_table = NULL; |
452 | } |
453 | |
454 | /* |
455 | * insert an rbio into the stripe cache. It |
456 | * must have already been prepared by calling |
457 | * cache_rbio_pages |
458 | * |
459 | * If this rbio was already cached, it gets |
460 | * moved to the front of the lru. |
461 | * |
462 | * If the size of the rbio cache is too big, we |
463 | * prune an item. |
464 | */ |
465 | static void cache_rbio(struct btrfs_raid_bio *rbio) |
466 | { |
467 | struct btrfs_stripe_hash_table *table; |
468 | |
469 | if (!test_bit(RBIO_CACHE_READY_BIT, &rbio->flags)) |
470 | return; |
471 | |
472 | table = rbio->bioc->fs_info->stripe_hash_table; |
473 | |
474 | spin_lock(lock: &table->cache_lock); |
475 | spin_lock(lock: &rbio->bio_list_lock); |
476 | |
477 | /* bump our ref if we were not in the list before */ |
478 | if (!test_and_set_bit(RBIO_CACHE_BIT, addr: &rbio->flags)) |
479 | refcount_inc(r: &rbio->refs); |
480 | |
481 | if (!list_empty(head: &rbio->stripe_cache)){ |
482 | list_move(list: &rbio->stripe_cache, head: &table->stripe_cache); |
483 | } else { |
484 | list_add(new: &rbio->stripe_cache, head: &table->stripe_cache); |
485 | table->cache_size += 1; |
486 | } |
487 | |
488 | spin_unlock(lock: &rbio->bio_list_lock); |
489 | |
490 | if (table->cache_size > RBIO_CACHE_SIZE) { |
491 | struct btrfs_raid_bio *found; |
492 | |
493 | found = list_entry(table->stripe_cache.prev, |
494 | struct btrfs_raid_bio, |
495 | stripe_cache); |
496 | |
497 | if (found != rbio) |
498 | __remove_rbio_from_cache(rbio: found); |
499 | } |
500 | |
501 | spin_unlock(lock: &table->cache_lock); |
502 | } |
503 | |
504 | /* |
505 | * helper function to run the xor_blocks api. It is only |
506 | * able to do MAX_XOR_BLOCKS at a time, so we need to |
507 | * loop through. |
508 | */ |
509 | static void run_xor(void **pages, int src_cnt, ssize_t len) |
510 | { |
511 | int src_off = 0; |
512 | int xor_src_cnt = 0; |
513 | void *dest = pages[src_cnt]; |
514 | |
515 | while(src_cnt > 0) { |
516 | xor_src_cnt = min(src_cnt, MAX_XOR_BLOCKS); |
517 | xor_blocks(count: xor_src_cnt, bytes: len, dest, srcs: pages + src_off); |
518 | |
519 | src_cnt -= xor_src_cnt; |
520 | src_off += xor_src_cnt; |
521 | } |
522 | } |
523 | |
524 | /* |
525 | * Returns true if the bio list inside this rbio covers an entire stripe (no |
526 | * rmw required). |
527 | */ |
528 | static int rbio_is_full(struct btrfs_raid_bio *rbio) |
529 | { |
530 | unsigned long size = rbio->bio_list_bytes; |
531 | int ret = 1; |
532 | |
533 | spin_lock(lock: &rbio->bio_list_lock); |
534 | if (size != rbio->nr_data * BTRFS_STRIPE_LEN) |
535 | ret = 0; |
536 | BUG_ON(size > rbio->nr_data * BTRFS_STRIPE_LEN); |
537 | spin_unlock(lock: &rbio->bio_list_lock); |
538 | |
539 | return ret; |
540 | } |
541 | |
542 | /* |
543 | * returns 1 if it is safe to merge two rbios together. |
544 | * The merging is safe if the two rbios correspond to |
545 | * the same stripe and if they are both going in the same |
546 | * direction (read vs write), and if neither one is |
547 | * locked for final IO |
548 | * |
549 | * The caller is responsible for locking such that |
550 | * rmw_locked is safe to test |
551 | */ |
552 | static int rbio_can_merge(struct btrfs_raid_bio *last, |
553 | struct btrfs_raid_bio *cur) |
554 | { |
555 | if (test_bit(RBIO_RMW_LOCKED_BIT, &last->flags) || |
556 | test_bit(RBIO_RMW_LOCKED_BIT, &cur->flags)) |
557 | return 0; |
558 | |
559 | /* |
560 | * we can't merge with cached rbios, since the |
561 | * idea is that when we merge the destination |
562 | * rbio is going to run our IO for us. We can |
563 | * steal from cached rbios though, other functions |
564 | * handle that. |
565 | */ |
566 | if (test_bit(RBIO_CACHE_BIT, &last->flags) || |
567 | test_bit(RBIO_CACHE_BIT, &cur->flags)) |
568 | return 0; |
569 | |
570 | if (last->bioc->full_stripe_logical != cur->bioc->full_stripe_logical) |
571 | return 0; |
572 | |
573 | /* we can't merge with different operations */ |
574 | if (last->operation != cur->operation) |
575 | return 0; |
576 | /* |
577 | * We've need read the full stripe from the drive. |
578 | * check and repair the parity and write the new results. |
579 | * |
580 | * We're not allowed to add any new bios to the |
581 | * bio list here, anyone else that wants to |
582 | * change this stripe needs to do their own rmw. |
583 | */ |
584 | if (last->operation == BTRFS_RBIO_PARITY_SCRUB) |
585 | return 0; |
586 | |
587 | if (last->operation == BTRFS_RBIO_READ_REBUILD) |
588 | return 0; |
589 | |
590 | return 1; |
591 | } |
592 | |
593 | static unsigned int rbio_stripe_sector_index(const struct btrfs_raid_bio *rbio, |
594 | unsigned int stripe_nr, |
595 | unsigned int sector_nr) |
596 | { |
597 | ASSERT(stripe_nr < rbio->real_stripes); |
598 | ASSERT(sector_nr < rbio->stripe_nsectors); |
599 | |
600 | return stripe_nr * rbio->stripe_nsectors + sector_nr; |
601 | } |
602 | |
603 | /* Return a sector from rbio->stripe_sectors, not from the bio list */ |
604 | static struct sector_ptr *rbio_stripe_sector(const struct btrfs_raid_bio *rbio, |
605 | unsigned int stripe_nr, |
606 | unsigned int sector_nr) |
607 | { |
608 | return &rbio->stripe_sectors[rbio_stripe_sector_index(rbio, stripe_nr, |
609 | sector_nr)]; |
610 | } |
611 | |
612 | /* Grab a sector inside P stripe */ |
613 | static struct sector_ptr *rbio_pstripe_sector(const struct btrfs_raid_bio *rbio, |
614 | unsigned int sector_nr) |
615 | { |
616 | return rbio_stripe_sector(rbio, stripe_nr: rbio->nr_data, sector_nr); |
617 | } |
618 | |
619 | /* Grab a sector inside Q stripe, return NULL if not RAID6 */ |
620 | static struct sector_ptr *rbio_qstripe_sector(const struct btrfs_raid_bio *rbio, |
621 | unsigned int sector_nr) |
622 | { |
623 | if (rbio->nr_data + 1 == rbio->real_stripes) |
624 | return NULL; |
625 | return rbio_stripe_sector(rbio, stripe_nr: rbio->nr_data + 1, sector_nr); |
626 | } |
627 | |
628 | /* |
629 | * The first stripe in the table for a logical address |
630 | * has the lock. rbios are added in one of three ways: |
631 | * |
632 | * 1) Nobody has the stripe locked yet. The rbio is given |
633 | * the lock and 0 is returned. The caller must start the IO |
634 | * themselves. |
635 | * |
636 | * 2) Someone has the stripe locked, but we're able to merge |
637 | * with the lock owner. The rbio is freed and the IO will |
638 | * start automatically along with the existing rbio. 1 is returned. |
639 | * |
640 | * 3) Someone has the stripe locked, but we're not able to merge. |
641 | * The rbio is added to the lock owner's plug list, or merged into |
642 | * an rbio already on the plug list. When the lock owner unlocks, |
643 | * the next rbio on the list is run and the IO is started automatically. |
644 | * 1 is returned |
645 | * |
646 | * If we return 0, the caller still owns the rbio and must continue with |
647 | * IO submission. If we return 1, the caller must assume the rbio has |
648 | * already been freed. |
649 | */ |
650 | static noinline int lock_stripe_add(struct btrfs_raid_bio *rbio) |
651 | { |
652 | struct btrfs_stripe_hash *h; |
653 | struct btrfs_raid_bio *cur; |
654 | struct btrfs_raid_bio *pending; |
655 | struct btrfs_raid_bio *freeit = NULL; |
656 | struct btrfs_raid_bio *cache_drop = NULL; |
657 | int ret = 0; |
658 | |
659 | h = rbio->bioc->fs_info->stripe_hash_table->table + rbio_bucket(rbio); |
660 | |
661 | spin_lock(lock: &h->lock); |
662 | list_for_each_entry(cur, &h->hash_list, hash_list) { |
663 | if (cur->bioc->full_stripe_logical != rbio->bioc->full_stripe_logical) |
664 | continue; |
665 | |
666 | spin_lock(lock: &cur->bio_list_lock); |
667 | |
668 | /* Can we steal this cached rbio's pages? */ |
669 | if (bio_list_empty(bl: &cur->bio_list) && |
670 | list_empty(head: &cur->plug_list) && |
671 | test_bit(RBIO_CACHE_BIT, &cur->flags) && |
672 | !test_bit(RBIO_RMW_LOCKED_BIT, &cur->flags)) { |
673 | list_del_init(entry: &cur->hash_list); |
674 | refcount_dec(r: &cur->refs); |
675 | |
676 | steal_rbio(src: cur, dest: rbio); |
677 | cache_drop = cur; |
678 | spin_unlock(lock: &cur->bio_list_lock); |
679 | |
680 | goto lockit; |
681 | } |
682 | |
683 | /* Can we merge into the lock owner? */ |
684 | if (rbio_can_merge(last: cur, cur: rbio)) { |
685 | merge_rbio(dest: cur, victim: rbio); |
686 | spin_unlock(lock: &cur->bio_list_lock); |
687 | freeit = rbio; |
688 | ret = 1; |
689 | goto out; |
690 | } |
691 | |
692 | |
693 | /* |
694 | * We couldn't merge with the running rbio, see if we can merge |
695 | * with the pending ones. We don't have to check for rmw_locked |
696 | * because there is no way they are inside finish_rmw right now |
697 | */ |
698 | list_for_each_entry(pending, &cur->plug_list, plug_list) { |
699 | if (rbio_can_merge(last: pending, cur: rbio)) { |
700 | merge_rbio(dest: pending, victim: rbio); |
701 | spin_unlock(lock: &cur->bio_list_lock); |
702 | freeit = rbio; |
703 | ret = 1; |
704 | goto out; |
705 | } |
706 | } |
707 | |
708 | /* |
709 | * No merging, put us on the tail of the plug list, our rbio |
710 | * will be started with the currently running rbio unlocks |
711 | */ |
712 | list_add_tail(new: &rbio->plug_list, head: &cur->plug_list); |
713 | spin_unlock(lock: &cur->bio_list_lock); |
714 | ret = 1; |
715 | goto out; |
716 | } |
717 | lockit: |
718 | refcount_inc(r: &rbio->refs); |
719 | list_add(new: &rbio->hash_list, head: &h->hash_list); |
720 | out: |
721 | spin_unlock(lock: &h->lock); |
722 | if (cache_drop) |
723 | remove_rbio_from_cache(rbio: cache_drop); |
724 | if (freeit) |
725 | free_raid_bio(rbio: freeit); |
726 | return ret; |
727 | } |
728 | |
729 | static void recover_rbio_work_locked(struct work_struct *work); |
730 | |
731 | /* |
732 | * called as rmw or parity rebuild is completed. If the plug list has more |
733 | * rbios waiting for this stripe, the next one on the list will be started |
734 | */ |
735 | static noinline void unlock_stripe(struct btrfs_raid_bio *rbio) |
736 | { |
737 | int bucket; |
738 | struct btrfs_stripe_hash *h; |
739 | int keep_cache = 0; |
740 | |
741 | bucket = rbio_bucket(rbio); |
742 | h = rbio->bioc->fs_info->stripe_hash_table->table + bucket; |
743 | |
744 | if (list_empty(head: &rbio->plug_list)) |
745 | cache_rbio(rbio); |
746 | |
747 | spin_lock(lock: &h->lock); |
748 | spin_lock(lock: &rbio->bio_list_lock); |
749 | |
750 | if (!list_empty(head: &rbio->hash_list)) { |
751 | /* |
752 | * if we're still cached and there is no other IO |
753 | * to perform, just leave this rbio here for others |
754 | * to steal from later |
755 | */ |
756 | if (list_empty(head: &rbio->plug_list) && |
757 | test_bit(RBIO_CACHE_BIT, &rbio->flags)) { |
758 | keep_cache = 1; |
759 | clear_bit(RBIO_RMW_LOCKED_BIT, addr: &rbio->flags); |
760 | BUG_ON(!bio_list_empty(&rbio->bio_list)); |
761 | goto done; |
762 | } |
763 | |
764 | list_del_init(entry: &rbio->hash_list); |
765 | refcount_dec(r: &rbio->refs); |
766 | |
767 | /* |
768 | * we use the plug list to hold all the rbios |
769 | * waiting for the chance to lock this stripe. |
770 | * hand the lock over to one of them. |
771 | */ |
772 | if (!list_empty(head: &rbio->plug_list)) { |
773 | struct btrfs_raid_bio *next; |
774 | struct list_head *head = rbio->plug_list.next; |
775 | |
776 | next = list_entry(head, struct btrfs_raid_bio, |
777 | plug_list); |
778 | |
779 | list_del_init(entry: &rbio->plug_list); |
780 | |
781 | list_add(new: &next->hash_list, head: &h->hash_list); |
782 | refcount_inc(r: &next->refs); |
783 | spin_unlock(lock: &rbio->bio_list_lock); |
784 | spin_unlock(lock: &h->lock); |
785 | |
786 | if (next->operation == BTRFS_RBIO_READ_REBUILD) { |
787 | start_async_work(rbio: next, work_func: recover_rbio_work_locked); |
788 | } else if (next->operation == BTRFS_RBIO_WRITE) { |
789 | steal_rbio(src: rbio, dest: next); |
790 | start_async_work(rbio: next, work_func: rmw_rbio_work_locked); |
791 | } else if (next->operation == BTRFS_RBIO_PARITY_SCRUB) { |
792 | steal_rbio(src: rbio, dest: next); |
793 | start_async_work(rbio: next, work_func: scrub_rbio_work_locked); |
794 | } |
795 | |
796 | goto done_nolock; |
797 | } |
798 | } |
799 | done: |
800 | spin_unlock(lock: &rbio->bio_list_lock); |
801 | spin_unlock(lock: &h->lock); |
802 | |
803 | done_nolock: |
804 | if (!keep_cache) |
805 | remove_rbio_from_cache(rbio); |
806 | } |
807 | |
808 | static void rbio_endio_bio_list(struct bio *cur, blk_status_t err) |
809 | { |
810 | struct bio *next; |
811 | |
812 | while (cur) { |
813 | next = cur->bi_next; |
814 | cur->bi_next = NULL; |
815 | cur->bi_status = err; |
816 | bio_endio(cur); |
817 | cur = next; |
818 | } |
819 | } |
820 | |
821 | /* |
822 | * this frees the rbio and runs through all the bios in the |
823 | * bio_list and calls end_io on them |
824 | */ |
825 | static void rbio_orig_end_io(struct btrfs_raid_bio *rbio, blk_status_t err) |
826 | { |
827 | struct bio *cur = bio_list_get(bl: &rbio->bio_list); |
828 | struct bio *; |
829 | |
830 | kfree(objp: rbio->csum_buf); |
831 | bitmap_free(bitmap: rbio->csum_bitmap); |
832 | rbio->csum_buf = NULL; |
833 | rbio->csum_bitmap = NULL; |
834 | |
835 | /* |
836 | * Clear the data bitmap, as the rbio may be cached for later usage. |
837 | * do this before before unlock_stripe() so there will be no new bio |
838 | * for this bio. |
839 | */ |
840 | bitmap_clear(map: &rbio->dbitmap, start: 0, nbits: rbio->stripe_nsectors); |
841 | |
842 | /* |
843 | * At this moment, rbio->bio_list is empty, however since rbio does not |
844 | * always have RBIO_RMW_LOCKED_BIT set and rbio is still linked on the |
845 | * hash list, rbio may be merged with others so that rbio->bio_list |
846 | * becomes non-empty. |
847 | * Once unlock_stripe() is done, rbio->bio_list will not be updated any |
848 | * more and we can call bio_endio() on all queued bios. |
849 | */ |
850 | unlock_stripe(rbio); |
851 | extra = bio_list_get(bl: &rbio->bio_list); |
852 | free_raid_bio(rbio); |
853 | |
854 | rbio_endio_bio_list(cur, err); |
855 | if (extra) |
856 | rbio_endio_bio_list(cur: extra, err); |
857 | } |
858 | |
859 | /* |
860 | * Get a sector pointer specified by its @stripe_nr and @sector_nr. |
861 | * |
862 | * @rbio: The raid bio |
863 | * @stripe_nr: Stripe number, valid range [0, real_stripe) |
864 | * @sector_nr: Sector number inside the stripe, |
865 | * valid range [0, stripe_nsectors) |
866 | * @bio_list_only: Whether to use sectors inside the bio list only. |
867 | * |
868 | * The read/modify/write code wants to reuse the original bio page as much |
869 | * as possible, and only use stripe_sectors as fallback. |
870 | */ |
871 | static struct sector_ptr *sector_in_rbio(struct btrfs_raid_bio *rbio, |
872 | int stripe_nr, int sector_nr, |
873 | bool bio_list_only) |
874 | { |
875 | struct sector_ptr *sector; |
876 | int index; |
877 | |
878 | ASSERT(stripe_nr >= 0 && stripe_nr < rbio->real_stripes); |
879 | ASSERT(sector_nr >= 0 && sector_nr < rbio->stripe_nsectors); |
880 | |
881 | index = stripe_nr * rbio->stripe_nsectors + sector_nr; |
882 | ASSERT(index >= 0 && index < rbio->nr_sectors); |
883 | |
884 | spin_lock(lock: &rbio->bio_list_lock); |
885 | sector = &rbio->bio_sectors[index]; |
886 | if (sector->page || bio_list_only) { |
887 | /* Don't return sector without a valid page pointer */ |
888 | if (!sector->page) |
889 | sector = NULL; |
890 | spin_unlock(lock: &rbio->bio_list_lock); |
891 | return sector; |
892 | } |
893 | spin_unlock(lock: &rbio->bio_list_lock); |
894 | |
895 | return &rbio->stripe_sectors[index]; |
896 | } |
897 | |
898 | /* |
899 | * allocation and initial setup for the btrfs_raid_bio. Not |
900 | * this does not allocate any pages for rbio->pages. |
901 | */ |
902 | static struct btrfs_raid_bio *alloc_rbio(struct btrfs_fs_info *fs_info, |
903 | struct btrfs_io_context *bioc) |
904 | { |
905 | const unsigned int real_stripes = bioc->num_stripes - bioc->replace_nr_stripes; |
906 | const unsigned int stripe_npages = BTRFS_STRIPE_LEN >> PAGE_SHIFT; |
907 | const unsigned int num_pages = stripe_npages * real_stripes; |
908 | const unsigned int stripe_nsectors = |
909 | BTRFS_STRIPE_LEN >> fs_info->sectorsize_bits; |
910 | const unsigned int num_sectors = stripe_nsectors * real_stripes; |
911 | struct btrfs_raid_bio *rbio; |
912 | |
913 | /* PAGE_SIZE must also be aligned to sectorsize for subpage support */ |
914 | ASSERT(IS_ALIGNED(PAGE_SIZE, fs_info->sectorsize)); |
915 | /* |
916 | * Our current stripe len should be fixed to 64k thus stripe_nsectors |
917 | * (at most 16) should be no larger than BITS_PER_LONG. |
918 | */ |
919 | ASSERT(stripe_nsectors <= BITS_PER_LONG); |
920 | |
921 | rbio = kzalloc(size: sizeof(*rbio), GFP_NOFS); |
922 | if (!rbio) |
923 | return ERR_PTR(error: -ENOMEM); |
924 | rbio->stripe_pages = kcalloc(n: num_pages, size: sizeof(struct page *), |
925 | GFP_NOFS); |
926 | rbio->bio_sectors = kcalloc(n: num_sectors, size: sizeof(struct sector_ptr), |
927 | GFP_NOFS); |
928 | rbio->stripe_sectors = kcalloc(n: num_sectors, size: sizeof(struct sector_ptr), |
929 | GFP_NOFS); |
930 | rbio->finish_pointers = kcalloc(n: real_stripes, size: sizeof(void *), GFP_NOFS); |
931 | rbio->error_bitmap = bitmap_zalloc(nbits: num_sectors, GFP_NOFS); |
932 | |
933 | if (!rbio->stripe_pages || !rbio->bio_sectors || !rbio->stripe_sectors || |
934 | !rbio->finish_pointers || !rbio->error_bitmap) { |
935 | free_raid_bio_pointers(rbio); |
936 | kfree(objp: rbio); |
937 | return ERR_PTR(error: -ENOMEM); |
938 | } |
939 | |
940 | bio_list_init(bl: &rbio->bio_list); |
941 | init_waitqueue_head(&rbio->io_wait); |
942 | INIT_LIST_HEAD(list: &rbio->plug_list); |
943 | spin_lock_init(&rbio->bio_list_lock); |
944 | INIT_LIST_HEAD(list: &rbio->stripe_cache); |
945 | INIT_LIST_HEAD(list: &rbio->hash_list); |
946 | btrfs_get_bioc(bioc); |
947 | rbio->bioc = bioc; |
948 | rbio->nr_pages = num_pages; |
949 | rbio->nr_sectors = num_sectors; |
950 | rbio->real_stripes = real_stripes; |
951 | rbio->stripe_npages = stripe_npages; |
952 | rbio->stripe_nsectors = stripe_nsectors; |
953 | refcount_set(r: &rbio->refs, n: 1); |
954 | atomic_set(v: &rbio->stripes_pending, i: 0); |
955 | |
956 | ASSERT(btrfs_nr_parity_stripes(bioc->map_type)); |
957 | rbio->nr_data = real_stripes - btrfs_nr_parity_stripes(type: bioc->map_type); |
958 | |
959 | return rbio; |
960 | } |
961 | |
962 | /* allocate pages for all the stripes in the bio, including parity */ |
963 | static int alloc_rbio_pages(struct btrfs_raid_bio *rbio) |
964 | { |
965 | int ret; |
966 | |
967 | ret = btrfs_alloc_page_array(nr_pages: rbio->nr_pages, page_array: rbio->stripe_pages); |
968 | if (ret < 0) |
969 | return ret; |
970 | /* Mapping all sectors */ |
971 | index_stripe_sectors(rbio); |
972 | return 0; |
973 | } |
974 | |
975 | /* only allocate pages for p/q stripes */ |
976 | static int alloc_rbio_parity_pages(struct btrfs_raid_bio *rbio) |
977 | { |
978 | const int data_pages = rbio->nr_data * rbio->stripe_npages; |
979 | int ret; |
980 | |
981 | ret = btrfs_alloc_page_array(nr_pages: rbio->nr_pages - data_pages, |
982 | page_array: rbio->stripe_pages + data_pages); |
983 | if (ret < 0) |
984 | return ret; |
985 | |
986 | index_stripe_sectors(rbio); |
987 | return 0; |
988 | } |
989 | |
990 | /* |
991 | * Return the total number of errors found in the vertical stripe of @sector_nr. |
992 | * |
993 | * @faila and @failb will also be updated to the first and second stripe |
994 | * number of the errors. |
995 | */ |
996 | static int get_rbio_veritical_errors(struct btrfs_raid_bio *rbio, int sector_nr, |
997 | int *faila, int *failb) |
998 | { |
999 | int stripe_nr; |
1000 | int found_errors = 0; |
1001 | |
1002 | if (faila || failb) { |
1003 | /* |
1004 | * Both @faila and @failb should be valid pointers if any of |
1005 | * them is specified. |
1006 | */ |
1007 | ASSERT(faila && failb); |
1008 | *faila = -1; |
1009 | *failb = -1; |
1010 | } |
1011 | |
1012 | for (stripe_nr = 0; stripe_nr < rbio->real_stripes; stripe_nr++) { |
1013 | int total_sector_nr = stripe_nr * rbio->stripe_nsectors + sector_nr; |
1014 | |
1015 | if (test_bit(total_sector_nr, rbio->error_bitmap)) { |
1016 | found_errors++; |
1017 | if (faila) { |
1018 | /* Update faila and failb. */ |
1019 | if (*faila < 0) |
1020 | *faila = stripe_nr; |
1021 | else if (*failb < 0) |
1022 | *failb = stripe_nr; |
1023 | } |
1024 | } |
1025 | } |
1026 | return found_errors; |
1027 | } |
1028 | |
1029 | /* |
1030 | * Add a single sector @sector into our list of bios for IO. |
1031 | * |
1032 | * Return 0 if everything went well. |
1033 | * Return <0 for error. |
1034 | */ |
1035 | static int rbio_add_io_sector(struct btrfs_raid_bio *rbio, |
1036 | struct bio_list *bio_list, |
1037 | struct sector_ptr *sector, |
1038 | unsigned int stripe_nr, |
1039 | unsigned int sector_nr, |
1040 | enum req_op op) |
1041 | { |
1042 | const u32 sectorsize = rbio->bioc->fs_info->sectorsize; |
1043 | struct bio *last = bio_list->tail; |
1044 | int ret; |
1045 | struct bio *bio; |
1046 | struct btrfs_io_stripe *stripe; |
1047 | u64 disk_start; |
1048 | |
1049 | /* |
1050 | * Note: here stripe_nr has taken device replace into consideration, |
1051 | * thus it can be larger than rbio->real_stripe. |
1052 | * So here we check against bioc->num_stripes, not rbio->real_stripes. |
1053 | */ |
1054 | ASSERT(stripe_nr >= 0 && stripe_nr < rbio->bioc->num_stripes); |
1055 | ASSERT(sector_nr >= 0 && sector_nr < rbio->stripe_nsectors); |
1056 | ASSERT(sector->page); |
1057 | |
1058 | stripe = &rbio->bioc->stripes[stripe_nr]; |
1059 | disk_start = stripe->physical + sector_nr * sectorsize; |
1060 | |
1061 | /* if the device is missing, just fail this stripe */ |
1062 | if (!stripe->dev->bdev) { |
1063 | int found_errors; |
1064 | |
1065 | set_bit(nr: stripe_nr * rbio->stripe_nsectors + sector_nr, |
1066 | addr: rbio->error_bitmap); |
1067 | |
1068 | /* Check if we have reached tolerance early. */ |
1069 | found_errors = get_rbio_veritical_errors(rbio, sector_nr, |
1070 | NULL, NULL); |
1071 | if (found_errors > rbio->bioc->max_errors) |
1072 | return -EIO; |
1073 | return 0; |
1074 | } |
1075 | |
1076 | /* see if we can add this page onto our existing bio */ |
1077 | if (last) { |
1078 | u64 last_end = last->bi_iter.bi_sector << SECTOR_SHIFT; |
1079 | last_end += last->bi_iter.bi_size; |
1080 | |
1081 | /* |
1082 | * we can't merge these if they are from different |
1083 | * devices or if they are not contiguous |
1084 | */ |
1085 | if (last_end == disk_start && !last->bi_status && |
1086 | last->bi_bdev == stripe->dev->bdev) { |
1087 | ret = bio_add_page(bio: last, page: sector->page, len: sectorsize, |
1088 | off: sector->pgoff); |
1089 | if (ret == sectorsize) |
1090 | return 0; |
1091 | } |
1092 | } |
1093 | |
1094 | /* put a new bio on the list */ |
1095 | bio = bio_alloc(bdev: stripe->dev->bdev, |
1096 | max(BTRFS_STRIPE_LEN >> PAGE_SHIFT, 1), |
1097 | opf: op, GFP_NOFS); |
1098 | bio->bi_iter.bi_sector = disk_start >> SECTOR_SHIFT; |
1099 | bio->bi_private = rbio; |
1100 | |
1101 | __bio_add_page(bio, page: sector->page, len: sectorsize, off: sector->pgoff); |
1102 | bio_list_add(bl: bio_list, bio); |
1103 | return 0; |
1104 | } |
1105 | |
1106 | static void index_one_bio(struct btrfs_raid_bio *rbio, struct bio *bio) |
1107 | { |
1108 | const u32 sectorsize = rbio->bioc->fs_info->sectorsize; |
1109 | struct bio_vec bvec; |
1110 | struct bvec_iter iter; |
1111 | u32 offset = (bio->bi_iter.bi_sector << SECTOR_SHIFT) - |
1112 | rbio->bioc->full_stripe_logical; |
1113 | |
1114 | bio_for_each_segment(bvec, bio, iter) { |
1115 | u32 bvec_offset; |
1116 | |
1117 | for (bvec_offset = 0; bvec_offset < bvec.bv_len; |
1118 | bvec_offset += sectorsize, offset += sectorsize) { |
1119 | int index = offset / sectorsize; |
1120 | struct sector_ptr *sector = &rbio->bio_sectors[index]; |
1121 | |
1122 | sector->page = bvec.bv_page; |
1123 | sector->pgoff = bvec.bv_offset + bvec_offset; |
1124 | ASSERT(sector->pgoff < PAGE_SIZE); |
1125 | } |
1126 | } |
1127 | } |
1128 | |
1129 | /* |
1130 | * helper function to walk our bio list and populate the bio_pages array with |
1131 | * the result. This seems expensive, but it is faster than constantly |
1132 | * searching through the bio list as we setup the IO in finish_rmw or stripe |
1133 | * reconstruction. |
1134 | * |
1135 | * This must be called before you trust the answers from page_in_rbio |
1136 | */ |
1137 | static void index_rbio_pages(struct btrfs_raid_bio *rbio) |
1138 | { |
1139 | struct bio *bio; |
1140 | |
1141 | spin_lock(lock: &rbio->bio_list_lock); |
1142 | bio_list_for_each(bio, &rbio->bio_list) |
1143 | index_one_bio(rbio, bio); |
1144 | |
1145 | spin_unlock(lock: &rbio->bio_list_lock); |
1146 | } |
1147 | |
1148 | static void bio_get_trace_info(struct btrfs_raid_bio *rbio, struct bio *bio, |
1149 | struct raid56_bio_trace_info *trace_info) |
1150 | { |
1151 | const struct btrfs_io_context *bioc = rbio->bioc; |
1152 | int i; |
1153 | |
1154 | ASSERT(bioc); |
1155 | |
1156 | /* We rely on bio->bi_bdev to find the stripe number. */ |
1157 | if (!bio->bi_bdev) |
1158 | goto not_found; |
1159 | |
1160 | for (i = 0; i < bioc->num_stripes; i++) { |
1161 | if (bio->bi_bdev != bioc->stripes[i].dev->bdev) |
1162 | continue; |
1163 | trace_info->stripe_nr = i; |
1164 | trace_info->devid = bioc->stripes[i].dev->devid; |
1165 | trace_info->offset = (bio->bi_iter.bi_sector << SECTOR_SHIFT) - |
1166 | bioc->stripes[i].physical; |
1167 | return; |
1168 | } |
1169 | |
1170 | not_found: |
1171 | trace_info->devid = -1; |
1172 | trace_info->offset = -1; |
1173 | trace_info->stripe_nr = -1; |
1174 | } |
1175 | |
1176 | static inline void bio_list_put(struct bio_list *bio_list) |
1177 | { |
1178 | struct bio *bio; |
1179 | |
1180 | while ((bio = bio_list_pop(bl: bio_list))) |
1181 | bio_put(bio); |
1182 | } |
1183 | |
1184 | /* Generate PQ for one vertical stripe. */ |
1185 | static void generate_pq_vertical(struct btrfs_raid_bio *rbio, int sectornr) |
1186 | { |
1187 | void **pointers = rbio->finish_pointers; |
1188 | const u32 sectorsize = rbio->bioc->fs_info->sectorsize; |
1189 | struct sector_ptr *sector; |
1190 | int stripe; |
1191 | const bool has_qstripe = rbio->bioc->map_type & BTRFS_BLOCK_GROUP_RAID6; |
1192 | |
1193 | /* First collect one sector from each data stripe */ |
1194 | for (stripe = 0; stripe < rbio->nr_data; stripe++) { |
1195 | sector = sector_in_rbio(rbio, stripe_nr: stripe, sector_nr: sectornr, bio_list_only: 0); |
1196 | pointers[stripe] = kmap_local_page(page: sector->page) + |
1197 | sector->pgoff; |
1198 | } |
1199 | |
1200 | /* Then add the parity stripe */ |
1201 | sector = rbio_pstripe_sector(rbio, sector_nr: sectornr); |
1202 | sector->uptodate = 1; |
1203 | pointers[stripe++] = kmap_local_page(page: sector->page) + sector->pgoff; |
1204 | |
1205 | if (has_qstripe) { |
1206 | /* |
1207 | * RAID6, add the qstripe and call the library function |
1208 | * to fill in our p/q |
1209 | */ |
1210 | sector = rbio_qstripe_sector(rbio, sector_nr: sectornr); |
1211 | sector->uptodate = 1; |
1212 | pointers[stripe++] = kmap_local_page(page: sector->page) + |
1213 | sector->pgoff; |
1214 | |
1215 | raid6_call.gen_syndrome(rbio->real_stripes, sectorsize, |
1216 | pointers); |
1217 | } else { |
1218 | /* raid5 */ |
1219 | memcpy(pointers[rbio->nr_data], pointers[0], sectorsize); |
1220 | run_xor(pages: pointers + 1, src_cnt: rbio->nr_data - 1, len: sectorsize); |
1221 | } |
1222 | for (stripe = stripe - 1; stripe >= 0; stripe--) |
1223 | kunmap_local(pointers[stripe]); |
1224 | } |
1225 | |
1226 | static int rmw_assemble_write_bios(struct btrfs_raid_bio *rbio, |
1227 | struct bio_list *bio_list) |
1228 | { |
1229 | /* The total sector number inside the full stripe. */ |
1230 | int total_sector_nr; |
1231 | int sectornr; |
1232 | int stripe; |
1233 | int ret; |
1234 | |
1235 | ASSERT(bio_list_size(bio_list) == 0); |
1236 | |
1237 | /* We should have at least one data sector. */ |
1238 | ASSERT(bitmap_weight(&rbio->dbitmap, rbio->stripe_nsectors)); |
1239 | |
1240 | /* |
1241 | * Reset errors, as we may have errors inherited from from degraded |
1242 | * write. |
1243 | */ |
1244 | bitmap_clear(map: rbio->error_bitmap, start: 0, nbits: rbio->nr_sectors); |
1245 | |
1246 | /* |
1247 | * Start assembly. Make bios for everything from the higher layers (the |
1248 | * bio_list in our rbio) and our P/Q. Ignore everything else. |
1249 | */ |
1250 | for (total_sector_nr = 0; total_sector_nr < rbio->nr_sectors; |
1251 | total_sector_nr++) { |
1252 | struct sector_ptr *sector; |
1253 | |
1254 | stripe = total_sector_nr / rbio->stripe_nsectors; |
1255 | sectornr = total_sector_nr % rbio->stripe_nsectors; |
1256 | |
1257 | /* This vertical stripe has no data, skip it. */ |
1258 | if (!test_bit(sectornr, &rbio->dbitmap)) |
1259 | continue; |
1260 | |
1261 | if (stripe < rbio->nr_data) { |
1262 | sector = sector_in_rbio(rbio, stripe_nr: stripe, sector_nr: sectornr, bio_list_only: 1); |
1263 | if (!sector) |
1264 | continue; |
1265 | } else { |
1266 | sector = rbio_stripe_sector(rbio, stripe_nr: stripe, sector_nr: sectornr); |
1267 | } |
1268 | |
1269 | ret = rbio_add_io_sector(rbio, bio_list, sector, stripe_nr: stripe, |
1270 | sector_nr: sectornr, op: REQ_OP_WRITE); |
1271 | if (ret) |
1272 | goto error; |
1273 | } |
1274 | |
1275 | if (likely(!rbio->bioc->replace_nr_stripes)) |
1276 | return 0; |
1277 | |
1278 | /* |
1279 | * Make a copy for the replace target device. |
1280 | * |
1281 | * Thus the source stripe number (in replace_stripe_src) should be valid. |
1282 | */ |
1283 | ASSERT(rbio->bioc->replace_stripe_src >= 0); |
1284 | |
1285 | for (total_sector_nr = 0; total_sector_nr < rbio->nr_sectors; |
1286 | total_sector_nr++) { |
1287 | struct sector_ptr *sector; |
1288 | |
1289 | stripe = total_sector_nr / rbio->stripe_nsectors; |
1290 | sectornr = total_sector_nr % rbio->stripe_nsectors; |
1291 | |
1292 | /* |
1293 | * For RAID56, there is only one device that can be replaced, |
1294 | * and replace_stripe_src[0] indicates the stripe number we |
1295 | * need to copy from. |
1296 | */ |
1297 | if (stripe != rbio->bioc->replace_stripe_src) { |
1298 | /* |
1299 | * We can skip the whole stripe completely, note |
1300 | * total_sector_nr will be increased by one anyway. |
1301 | */ |
1302 | ASSERT(sectornr == 0); |
1303 | total_sector_nr += rbio->stripe_nsectors - 1; |
1304 | continue; |
1305 | } |
1306 | |
1307 | /* This vertical stripe has no data, skip it. */ |
1308 | if (!test_bit(sectornr, &rbio->dbitmap)) |
1309 | continue; |
1310 | |
1311 | if (stripe < rbio->nr_data) { |
1312 | sector = sector_in_rbio(rbio, stripe_nr: stripe, sector_nr: sectornr, bio_list_only: 1); |
1313 | if (!sector) |
1314 | continue; |
1315 | } else { |
1316 | sector = rbio_stripe_sector(rbio, stripe_nr: stripe, sector_nr: sectornr); |
1317 | } |
1318 | |
1319 | ret = rbio_add_io_sector(rbio, bio_list, sector, |
1320 | stripe_nr: rbio->real_stripes, |
1321 | sector_nr: sectornr, op: REQ_OP_WRITE); |
1322 | if (ret) |
1323 | goto error; |
1324 | } |
1325 | |
1326 | return 0; |
1327 | error: |
1328 | bio_list_put(bio_list); |
1329 | return -EIO; |
1330 | } |
1331 | |
1332 | static void set_rbio_range_error(struct btrfs_raid_bio *rbio, struct bio *bio) |
1333 | { |
1334 | struct btrfs_fs_info *fs_info = rbio->bioc->fs_info; |
1335 | u32 offset = (bio->bi_iter.bi_sector << SECTOR_SHIFT) - |
1336 | rbio->bioc->full_stripe_logical; |
1337 | int total_nr_sector = offset >> fs_info->sectorsize_bits; |
1338 | |
1339 | ASSERT(total_nr_sector < rbio->nr_data * rbio->stripe_nsectors); |
1340 | |
1341 | bitmap_set(map: rbio->error_bitmap, start: total_nr_sector, |
1342 | nbits: bio->bi_iter.bi_size >> fs_info->sectorsize_bits); |
1343 | |
1344 | /* |
1345 | * Special handling for raid56_alloc_missing_rbio() used by |
1346 | * scrub/replace. Unlike call path in raid56_parity_recover(), they |
1347 | * pass an empty bio here. Thus we have to find out the missing device |
1348 | * and mark the stripe error instead. |
1349 | */ |
1350 | if (bio->bi_iter.bi_size == 0) { |
1351 | bool found_missing = false; |
1352 | int stripe_nr; |
1353 | |
1354 | for (stripe_nr = 0; stripe_nr < rbio->real_stripes; stripe_nr++) { |
1355 | if (!rbio->bioc->stripes[stripe_nr].dev->bdev) { |
1356 | found_missing = true; |
1357 | bitmap_set(map: rbio->error_bitmap, |
1358 | start: stripe_nr * rbio->stripe_nsectors, |
1359 | nbits: rbio->stripe_nsectors); |
1360 | } |
1361 | } |
1362 | ASSERT(found_missing); |
1363 | } |
1364 | } |
1365 | |
1366 | /* |
1367 | * For subpage case, we can no longer set page Up-to-date directly for |
1368 | * stripe_pages[], thus we need to locate the sector. |
1369 | */ |
1370 | static struct sector_ptr *find_stripe_sector(struct btrfs_raid_bio *rbio, |
1371 | struct page *page, |
1372 | unsigned int pgoff) |
1373 | { |
1374 | int i; |
1375 | |
1376 | for (i = 0; i < rbio->nr_sectors; i++) { |
1377 | struct sector_ptr *sector = &rbio->stripe_sectors[i]; |
1378 | |
1379 | if (sector->page == page && sector->pgoff == pgoff) |
1380 | return sector; |
1381 | } |
1382 | return NULL; |
1383 | } |
1384 | |
1385 | /* |
1386 | * this sets each page in the bio uptodate. It should only be used on private |
1387 | * rbio pages, nothing that comes in from the higher layers |
1388 | */ |
1389 | static void set_bio_pages_uptodate(struct btrfs_raid_bio *rbio, struct bio *bio) |
1390 | { |
1391 | const u32 sectorsize = rbio->bioc->fs_info->sectorsize; |
1392 | struct bio_vec *bvec; |
1393 | struct bvec_iter_all iter_all; |
1394 | |
1395 | ASSERT(!bio_flagged(bio, BIO_CLONED)); |
1396 | |
1397 | bio_for_each_segment_all(bvec, bio, iter_all) { |
1398 | struct sector_ptr *sector; |
1399 | int pgoff; |
1400 | |
1401 | for (pgoff = bvec->bv_offset; pgoff - bvec->bv_offset < bvec->bv_len; |
1402 | pgoff += sectorsize) { |
1403 | sector = find_stripe_sector(rbio, page: bvec->bv_page, pgoff); |
1404 | ASSERT(sector); |
1405 | if (sector) |
1406 | sector->uptodate = 1; |
1407 | } |
1408 | } |
1409 | } |
1410 | |
1411 | static int get_bio_sector_nr(struct btrfs_raid_bio *rbio, struct bio *bio) |
1412 | { |
1413 | struct bio_vec *bv = bio_first_bvec_all(bio); |
1414 | int i; |
1415 | |
1416 | for (i = 0; i < rbio->nr_sectors; i++) { |
1417 | struct sector_ptr *sector; |
1418 | |
1419 | sector = &rbio->stripe_sectors[i]; |
1420 | if (sector->page == bv->bv_page && sector->pgoff == bv->bv_offset) |
1421 | break; |
1422 | sector = &rbio->bio_sectors[i]; |
1423 | if (sector->page == bv->bv_page && sector->pgoff == bv->bv_offset) |
1424 | break; |
1425 | } |
1426 | ASSERT(i < rbio->nr_sectors); |
1427 | return i; |
1428 | } |
1429 | |
1430 | static void rbio_update_error_bitmap(struct btrfs_raid_bio *rbio, struct bio *bio) |
1431 | { |
1432 | int total_sector_nr = get_bio_sector_nr(rbio, bio); |
1433 | u32 bio_size = 0; |
1434 | struct bio_vec *bvec; |
1435 | int i; |
1436 | |
1437 | bio_for_each_bvec_all(bvec, bio, i) |
1438 | bio_size += bvec->bv_len; |
1439 | |
1440 | /* |
1441 | * Since we can have multiple bios touching the error_bitmap, we cannot |
1442 | * call bitmap_set() without protection. |
1443 | * |
1444 | * Instead use set_bit() for each bit, as set_bit() itself is atomic. |
1445 | */ |
1446 | for (i = total_sector_nr; i < total_sector_nr + |
1447 | (bio_size >> rbio->bioc->fs_info->sectorsize_bits); i++) |
1448 | set_bit(nr: i, addr: rbio->error_bitmap); |
1449 | } |
1450 | |
1451 | /* Verify the data sectors at read time. */ |
1452 | static void verify_bio_data_sectors(struct btrfs_raid_bio *rbio, |
1453 | struct bio *bio) |
1454 | { |
1455 | struct btrfs_fs_info *fs_info = rbio->bioc->fs_info; |
1456 | int total_sector_nr = get_bio_sector_nr(rbio, bio); |
1457 | struct bio_vec *bvec; |
1458 | struct bvec_iter_all iter_all; |
1459 | |
1460 | /* No data csum for the whole stripe, no need to verify. */ |
1461 | if (!rbio->csum_bitmap || !rbio->csum_buf) |
1462 | return; |
1463 | |
1464 | /* P/Q stripes, they have no data csum to verify against. */ |
1465 | if (total_sector_nr >= rbio->nr_data * rbio->stripe_nsectors) |
1466 | return; |
1467 | |
1468 | bio_for_each_segment_all(bvec, bio, iter_all) { |
1469 | int bv_offset; |
1470 | |
1471 | for (bv_offset = bvec->bv_offset; |
1472 | bv_offset < bvec->bv_offset + bvec->bv_len; |
1473 | bv_offset += fs_info->sectorsize, total_sector_nr++) { |
1474 | u8 csum_buf[BTRFS_CSUM_SIZE]; |
1475 | u8 *expected_csum = rbio->csum_buf + |
1476 | total_sector_nr * fs_info->csum_size; |
1477 | int ret; |
1478 | |
1479 | /* No csum for this sector, skip to the next sector. */ |
1480 | if (!test_bit(total_sector_nr, rbio->csum_bitmap)) |
1481 | continue; |
1482 | |
1483 | ret = btrfs_check_sector_csum(fs_info, page: bvec->bv_page, |
1484 | pgoff: bv_offset, csum: csum_buf, csum_expected: expected_csum); |
1485 | if (ret < 0) |
1486 | set_bit(nr: total_sector_nr, addr: rbio->error_bitmap); |
1487 | } |
1488 | } |
1489 | } |
1490 | |
1491 | static void raid_wait_read_end_io(struct bio *bio) |
1492 | { |
1493 | struct btrfs_raid_bio *rbio = bio->bi_private; |
1494 | |
1495 | if (bio->bi_status) { |
1496 | rbio_update_error_bitmap(rbio, bio); |
1497 | } else { |
1498 | set_bio_pages_uptodate(rbio, bio); |
1499 | verify_bio_data_sectors(rbio, bio); |
1500 | } |
1501 | |
1502 | bio_put(bio); |
1503 | if (atomic_dec_and_test(v: &rbio->stripes_pending)) |
1504 | wake_up(&rbio->io_wait); |
1505 | } |
1506 | |
1507 | static void submit_read_wait_bio_list(struct btrfs_raid_bio *rbio, |
1508 | struct bio_list *bio_list) |
1509 | { |
1510 | struct bio *bio; |
1511 | |
1512 | atomic_set(v: &rbio->stripes_pending, i: bio_list_size(bl: bio_list)); |
1513 | while ((bio = bio_list_pop(bl: bio_list))) { |
1514 | bio->bi_end_io = raid_wait_read_end_io; |
1515 | |
1516 | if (trace_raid56_read_enabled()) { |
1517 | struct raid56_bio_trace_info trace_info = { 0 }; |
1518 | |
1519 | bio_get_trace_info(rbio, bio, trace_info: &trace_info); |
1520 | trace_raid56_read(rbio, bio, trace_info: &trace_info); |
1521 | } |
1522 | submit_bio(bio); |
1523 | } |
1524 | |
1525 | wait_event(rbio->io_wait, atomic_read(&rbio->stripes_pending) == 0); |
1526 | } |
1527 | |
1528 | static int alloc_rbio_data_pages(struct btrfs_raid_bio *rbio) |
1529 | { |
1530 | const int data_pages = rbio->nr_data * rbio->stripe_npages; |
1531 | int ret; |
1532 | |
1533 | ret = btrfs_alloc_page_array(nr_pages: data_pages, page_array: rbio->stripe_pages); |
1534 | if (ret < 0) |
1535 | return ret; |
1536 | |
1537 | index_stripe_sectors(rbio); |
1538 | return 0; |
1539 | } |
1540 | |
1541 | /* |
1542 | * We use plugging call backs to collect full stripes. |
1543 | * Any time we get a partial stripe write while plugged |
1544 | * we collect it into a list. When the unplug comes down, |
1545 | * we sort the list by logical block number and merge |
1546 | * everything we can into the same rbios |
1547 | */ |
1548 | struct btrfs_plug_cb { |
1549 | struct blk_plug_cb cb; |
1550 | struct btrfs_fs_info *info; |
1551 | struct list_head rbio_list; |
1552 | struct work_struct work; |
1553 | }; |
1554 | |
1555 | /* |
1556 | * rbios on the plug list are sorted for easier merging. |
1557 | */ |
1558 | static int plug_cmp(void *priv, const struct list_head *a, |
1559 | const struct list_head *b) |
1560 | { |
1561 | const struct btrfs_raid_bio *ra = container_of(a, struct btrfs_raid_bio, |
1562 | plug_list); |
1563 | const struct btrfs_raid_bio *rb = container_of(b, struct btrfs_raid_bio, |
1564 | plug_list); |
1565 | u64 a_sector = ra->bio_list.head->bi_iter.bi_sector; |
1566 | u64 b_sector = rb->bio_list.head->bi_iter.bi_sector; |
1567 | |
1568 | if (a_sector < b_sector) |
1569 | return -1; |
1570 | if (a_sector > b_sector) |
1571 | return 1; |
1572 | return 0; |
1573 | } |
1574 | |
1575 | static void raid_unplug(struct blk_plug_cb *cb, bool from_schedule) |
1576 | { |
1577 | struct btrfs_plug_cb *plug = container_of(cb, struct btrfs_plug_cb, cb); |
1578 | struct btrfs_raid_bio *cur; |
1579 | struct btrfs_raid_bio *last = NULL; |
1580 | |
1581 | list_sort(NULL, head: &plug->rbio_list, cmp: plug_cmp); |
1582 | |
1583 | while (!list_empty(head: &plug->rbio_list)) { |
1584 | cur = list_entry(plug->rbio_list.next, |
1585 | struct btrfs_raid_bio, plug_list); |
1586 | list_del_init(entry: &cur->plug_list); |
1587 | |
1588 | if (rbio_is_full(rbio: cur)) { |
1589 | /* We have a full stripe, queue it down. */ |
1590 | start_async_work(rbio: cur, work_func: rmw_rbio_work); |
1591 | continue; |
1592 | } |
1593 | if (last) { |
1594 | if (rbio_can_merge(last, cur)) { |
1595 | merge_rbio(dest: last, victim: cur); |
1596 | free_raid_bio(rbio: cur); |
1597 | continue; |
1598 | } |
1599 | start_async_work(rbio: last, work_func: rmw_rbio_work); |
1600 | } |
1601 | last = cur; |
1602 | } |
1603 | if (last) |
1604 | start_async_work(rbio: last, work_func: rmw_rbio_work); |
1605 | kfree(objp: plug); |
1606 | } |
1607 | |
1608 | /* Add the original bio into rbio->bio_list, and update rbio::dbitmap. */ |
1609 | static void rbio_add_bio(struct btrfs_raid_bio *rbio, struct bio *orig_bio) |
1610 | { |
1611 | const struct btrfs_fs_info *fs_info = rbio->bioc->fs_info; |
1612 | const u64 orig_logical = orig_bio->bi_iter.bi_sector << SECTOR_SHIFT; |
1613 | const u64 full_stripe_start = rbio->bioc->full_stripe_logical; |
1614 | const u32 orig_len = orig_bio->bi_iter.bi_size; |
1615 | const u32 sectorsize = fs_info->sectorsize; |
1616 | u64 cur_logical; |
1617 | |
1618 | ASSERT(orig_logical >= full_stripe_start && |
1619 | orig_logical + orig_len <= full_stripe_start + |
1620 | rbio->nr_data * BTRFS_STRIPE_LEN); |
1621 | |
1622 | bio_list_add(bl: &rbio->bio_list, bio: orig_bio); |
1623 | rbio->bio_list_bytes += orig_bio->bi_iter.bi_size; |
1624 | |
1625 | /* Update the dbitmap. */ |
1626 | for (cur_logical = orig_logical; cur_logical < orig_logical + orig_len; |
1627 | cur_logical += sectorsize) { |
1628 | int bit = ((u32)(cur_logical - full_stripe_start) >> |
1629 | fs_info->sectorsize_bits) % rbio->stripe_nsectors; |
1630 | |
1631 | set_bit(nr: bit, addr: &rbio->dbitmap); |
1632 | } |
1633 | } |
1634 | |
1635 | /* |
1636 | * our main entry point for writes from the rest of the FS. |
1637 | */ |
1638 | void raid56_parity_write(struct bio *bio, struct btrfs_io_context *bioc) |
1639 | { |
1640 | struct btrfs_fs_info *fs_info = bioc->fs_info; |
1641 | struct btrfs_raid_bio *rbio; |
1642 | struct btrfs_plug_cb *plug = NULL; |
1643 | struct blk_plug_cb *cb; |
1644 | |
1645 | rbio = alloc_rbio(fs_info, bioc); |
1646 | if (IS_ERR(ptr: rbio)) { |
1647 | bio->bi_status = errno_to_blk_status(errno: PTR_ERR(ptr: rbio)); |
1648 | bio_endio(bio); |
1649 | return; |
1650 | } |
1651 | rbio->operation = BTRFS_RBIO_WRITE; |
1652 | rbio_add_bio(rbio, orig_bio: bio); |
1653 | |
1654 | /* |
1655 | * Don't plug on full rbios, just get them out the door |
1656 | * as quickly as we can |
1657 | */ |
1658 | if (!rbio_is_full(rbio)) { |
1659 | cb = blk_check_plugged(unplug: raid_unplug, data: fs_info, size: sizeof(*plug)); |
1660 | if (cb) { |
1661 | plug = container_of(cb, struct btrfs_plug_cb, cb); |
1662 | if (!plug->info) { |
1663 | plug->info = fs_info; |
1664 | INIT_LIST_HEAD(list: &plug->rbio_list); |
1665 | } |
1666 | list_add_tail(new: &rbio->plug_list, head: &plug->rbio_list); |
1667 | return; |
1668 | } |
1669 | } |
1670 | |
1671 | /* |
1672 | * Either we don't have any existing plug, or we're doing a full stripe, |
1673 | * queue the rmw work now. |
1674 | */ |
1675 | start_async_work(rbio, work_func: rmw_rbio_work); |
1676 | } |
1677 | |
1678 | static int verify_one_sector(struct btrfs_raid_bio *rbio, |
1679 | int stripe_nr, int sector_nr) |
1680 | { |
1681 | struct btrfs_fs_info *fs_info = rbio->bioc->fs_info; |
1682 | struct sector_ptr *sector; |
1683 | u8 csum_buf[BTRFS_CSUM_SIZE]; |
1684 | u8 *csum_expected; |
1685 | int ret; |
1686 | |
1687 | if (!rbio->csum_bitmap || !rbio->csum_buf) |
1688 | return 0; |
1689 | |
1690 | /* No way to verify P/Q as they are not covered by data csum. */ |
1691 | if (stripe_nr >= rbio->nr_data) |
1692 | return 0; |
1693 | /* |
1694 | * If we're rebuilding a read, we have to use pages from the |
1695 | * bio list if possible. |
1696 | */ |
1697 | if (rbio->operation == BTRFS_RBIO_READ_REBUILD) { |
1698 | sector = sector_in_rbio(rbio, stripe_nr, sector_nr, bio_list_only: 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, page: sector->page, pgoff: sector->pgoff, |
1709 | csum: csum_buf, csum_expected); |
1710 | return ret; |
1711 | } |
1712 | |
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 | */ |
1718 | static int recover_vertical(struct btrfs_raid_bio *rbio, int sector_nr, |
1719 | void **pointers, void **unmap_array) |
1720 | { |
1721 | struct btrfs_fs_info *fs_info = rbio->bioc->fs_info; |
1722 | struct sector_ptr *sector; |
1723 | const u32 sectorsize = fs_info->sectorsize; |
1724 | int found_errors; |
1725 | int faila; |
1726 | int failb; |
1727 | int stripe_nr; |
1728 | int ret = 0; |
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)) |
1736 | return 0; |
1737 | |
1738 | found_errors = get_rbio_veritical_errors(rbio, sector_nr, faila: &faila, |
1739 | failb: &failb); |
1740 | /* |
1741 | * No errors in the vertical stripe, skip it. Can happen for recovery |
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; |
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 | /* |
1758 | * If we're rebuilding a read, we have to use pages from the |
1759 | * bio list if possible. |
1760 | */ |
1761 | if (rbio->operation == BTRFS_RBIO_READ_REBUILD) { |
1762 | sector = sector_in_rbio(rbio, stripe_nr, sector_nr, bio_list_only: 0); |
1763 | } else { |
1764 | sector = rbio_stripe_sector(rbio, stripe_nr, sector_nr); |
1765 | } |
1766 | ASSERT(sector->page); |
1767 | pointers[stripe_nr] = kmap_local_page(page: sector->page) + |
1768 | sector->pgoff; |
1769 | unmap_array[stripe_nr] = pointers[stripe_nr]; |
1770 | } |
1771 | |
1772 | /* All raid6 handling here */ |
1773 | if (rbio->bioc->map_type & BTRFS_BLOCK_GROUP_RAID6) { |
1774 | /* Single failure, rebuild from parity raid5 style */ |
1775 | if (failb < 0) { |
1776 | if (faila == rbio->nr_data) |
1777 | /* |
1778 | * Just the P stripe has failed, without |
1779 | * a bad data or Q stripe. |
1780 | * We have nothing to do, just skip the |
1781 | * recovery for this stripe. |
1782 | */ |
1783 | goto cleanup; |
1784 | /* |
1785 | * a single failure in raid6 is rebuilt |
1786 | * in the pstripe code below |
1787 | */ |
1788 | goto pstripe; |
1789 | } |
1790 | |
1791 | /* |
1792 | * If the q stripe is failed, do a pstripe reconstruction from |
1793 | * the xors. |
1794 | * If both the q stripe and the P stripe are failed, we're |
1795 | * here due to a crc mismatch and we can't give them the |
1796 | * data they want. |
1797 | */ |
1798 | if (failb == rbio->real_stripes - 1) { |
1799 | if (faila == rbio->real_stripes - 2) |
1800 | /* |
1801 | * Only P and Q are corrupted. |
1802 | * We only care about data stripes recovery, |
1803 | * can skip this vertical stripe. |
1804 | */ |
1805 | goto cleanup; |
1806 | /* |
1807 | * Otherwise we have one bad data stripe and |
1808 | * a good P stripe. raid5! |
1809 | */ |
1810 | goto pstripe; |
1811 | } |
1812 | |
1813 | if (failb == rbio->real_stripes - 2) { |
1814 | raid6_datap_recov(rbio->real_stripes, sectorsize, |
1815 | faila, pointers); |
1816 | } else { |
1817 | raid6_2data_recov(rbio->real_stripes, sectorsize, |
1818 | faila, failb, pointers); |
1819 | } |
1820 | } else { |
1821 | void *p; |
1822 | |
1823 | /* Rebuild from P stripe here (raid5 or raid6). */ |
1824 | ASSERT(failb == -1); |
1825 | pstripe: |
1826 | /* Copy parity block into failed block to start with */ |
1827 | memcpy(pointers[faila], pointers[rbio->nr_data], sectorsize); |
1828 | |
1829 | /* Rearrange the pointer array */ |
1830 | p = pointers[faila]; |
1831 | for (stripe_nr = faila; stripe_nr < rbio->nr_data - 1; |
1832 | stripe_nr++) |
1833 | pointers[stripe_nr] = pointers[stripe_nr + 1]; |
1834 | pointers[rbio->nr_data - 1] = p; |
1835 | |
1836 | /* Xor in the rest */ |
1837 | run_xor(pages: pointers, src_cnt: rbio->nr_data - 1, len: sectorsize); |
1838 | |
1839 | } |
1840 | |
1841 | /* |
1842 | * No matter if this is a RMW or recovery, we should have all |
1843 | * failed sectors repaired in the vertical stripe, thus they are now |
1844 | * uptodate. |
1845 | * Especially if we determine to cache the rbio, we need to |
1846 | * have at least all data sectors uptodate. |
1847 | * |
1848 | * If possible, also check if the repaired sector matches its data |
1849 | * checksum. |
1850 | */ |
1851 | if (faila >= 0) { |
1852 | ret = verify_one_sector(rbio, stripe_nr: faila, sector_nr); |
1853 | if (ret < 0) |
1854 | goto cleanup; |
1855 | |
1856 | sector = rbio_stripe_sector(rbio, stripe_nr: faila, sector_nr); |
1857 | sector->uptodate = 1; |
1858 | } |
1859 | if (failb >= 0) { |
1860 | ret = verify_one_sector(rbio, stripe_nr: failb, sector_nr); |
1861 | if (ret < 0) |
1862 | goto cleanup; |
1863 | |
1864 | sector = rbio_stripe_sector(rbio, stripe_nr: failb, sector_nr); |
1865 | sector->uptodate = 1; |
1866 | } |
1867 | |
1868 | cleanup: |
1869 | for (stripe_nr = rbio->real_stripes - 1; stripe_nr >= 0; stripe_nr--) |
1870 | kunmap_local(unmap_array[stripe_nr]); |
1871 | return ret; |
1872 | } |
1873 | |
1874 | static int recover_sectors(struct btrfs_raid_bio *rbio) |
1875 | { |
1876 | void **pointers = NULL; |
1877 | void **unmap_array = NULL; |
1878 | int sectornr; |
1879 | int ret = 0; |
1880 | |
1881 | /* |
1882 | * @pointers array stores the pointer for each sector. |
1883 | * |
1884 | * @unmap_array stores copy of pointers that does not get reordered |
1885 | * during reconstruction so that kunmap_local works. |
1886 | */ |
1887 | pointers = kcalloc(n: rbio->real_stripes, size: sizeof(void *), GFP_NOFS); |
1888 | unmap_array = kcalloc(n: rbio->real_stripes, size: sizeof(void *), GFP_NOFS); |
1889 | if (!pointers || !unmap_array) { |
1890 | ret = -ENOMEM; |
1891 | goto out; |
1892 | } |
1893 | |
1894 | if (rbio->operation == BTRFS_RBIO_READ_REBUILD) { |
1895 | spin_lock(lock: &rbio->bio_list_lock); |
1896 | set_bit(RBIO_RMW_LOCKED_BIT, addr: &rbio->flags); |
1897 | spin_unlock(lock: &rbio->bio_list_lock); |
1898 | } |
1899 | |
1900 | index_rbio_pages(rbio); |
1901 | |
1902 | for (sectornr = 0; sectornr < rbio->stripe_nsectors; sectornr++) { |
1903 | ret = recover_vertical(rbio, sector_nr: sectornr, pointers, unmap_array); |
1904 | if (ret < 0) |
1905 | break; |
1906 | } |
1907 | |
1908 | out: |
1909 | kfree(objp: pointers); |
1910 | kfree(objp: unmap_array); |
1911 | return ret; |
1912 | } |
1913 | |
1914 | static void recover_rbio(struct btrfs_raid_bio *rbio) |
1915 | { |
1916 | struct bio_list bio_list = BIO_EMPTY_LIST; |
1917 | int total_sector_nr; |
1918 | int ret = 0; |
1919 | |
1920 | /* |
1921 | * Either we're doing recover for a read failure or degraded write, |
1922 | * caller should have set error bitmap correctly. |
1923 | */ |
1924 | ASSERT(bitmap_weight(rbio->error_bitmap, rbio->nr_sectors)); |
1925 | |
1926 | /* For recovery, we need to read all sectors including P/Q. */ |
1927 | ret = alloc_rbio_pages(rbio); |
1928 | if (ret < 0) |
1929 | goto out; |
1930 | |
1931 | index_rbio_pages(rbio); |
1932 | |
1933 | /* |
1934 | * Read everything that hasn't failed. However this time we will |
1935 | * not trust any cached sector. |
1936 | * As we may read out some stale data but higher layer is not reading |
1937 | * that stale part. |
1938 | * |
1939 | * So here we always re-read everything in recovery path. |
1940 | */ |
1941 | for (total_sector_nr = 0; total_sector_nr < rbio->nr_sectors; |
1942 | total_sector_nr++) { |
1943 | int stripe = total_sector_nr / rbio->stripe_nsectors; |
1944 | int sectornr = total_sector_nr % rbio->stripe_nsectors; |
1945 | struct sector_ptr *sector; |
1946 | |
1947 | /* |
1948 | * Skip the range which has error. It can be a range which is |
1949 | * marked error (for csum mismatch), or it can be a missing |
1950 | * device. |
1951 | */ |
1952 | if (!rbio->bioc->stripes[stripe].dev->bdev || |
1953 | test_bit(total_sector_nr, rbio->error_bitmap)) { |
1954 | /* |
1955 | * Also set the error bit for missing device, which |
1956 | * may not yet have its error bit set. |
1957 | */ |
1958 | set_bit(nr: total_sector_nr, addr: rbio->error_bitmap); |
1959 | continue; |
1960 | } |
1961 | |
1962 | sector = rbio_stripe_sector(rbio, stripe_nr: stripe, sector_nr: sectornr); |
1963 | ret = rbio_add_io_sector(rbio, bio_list: &bio_list, sector, stripe_nr: stripe, |
1964 | sector_nr: sectornr, op: REQ_OP_READ); |
1965 | if (ret < 0) { |
1966 | bio_list_put(bio_list: &bio_list); |
1967 | goto out; |
1968 | } |
1969 | } |
1970 | |
1971 | submit_read_wait_bio_list(rbio, bio_list: &bio_list); |
1972 | ret = recover_sectors(rbio); |
1973 | out: |
1974 | rbio_orig_end_io(rbio, err: errno_to_blk_status(errno: ret)); |
1975 | } |
1976 | |
1977 | static void recover_rbio_work(struct work_struct *work) |
1978 | { |
1979 | struct btrfs_raid_bio *rbio; |
1980 | |
1981 | rbio = container_of(work, struct btrfs_raid_bio, work); |
1982 | if (!lock_stripe_add(rbio)) |
1983 | recover_rbio(rbio); |
1984 | } |
1985 | |
1986 | static void recover_rbio_work_locked(struct work_struct *work) |
1987 | { |
1988 | recover_rbio(container_of(work, struct btrfs_raid_bio, work)); |
1989 | } |
1990 | |
1991 | static void (struct btrfs_raid_bio *rbio, int mirror_num) |
1992 | { |
1993 | bool found = false; |
1994 | int sector_nr; |
1995 | |
1996 | /* |
1997 | * This is for RAID6 extra recovery tries, thus mirror number should |
1998 | * be large than 2. |
1999 | * Mirror 1 means read from data stripes. Mirror 2 means rebuild using |
2000 | * RAID5 methods. |
2001 | */ |
2002 | ASSERT(mirror_num > 2); |
2003 | for (sector_nr = 0; sector_nr < rbio->stripe_nsectors; sector_nr++) { |
2004 | int found_errors; |
2005 | int faila; |
2006 | int failb; |
2007 | |
2008 | found_errors = get_rbio_veritical_errors(rbio, sector_nr, |
2009 | faila: &faila, failb: &failb); |
2010 | /* This vertical stripe doesn't have errors. */ |
2011 | if (!found_errors) |
2012 | continue; |
2013 | |
2014 | /* |
2015 | * If we found errors, there should be only one error marked |
2016 | * by previous set_rbio_range_error(). |
2017 | */ |
2018 | ASSERT(found_errors == 1); |
2019 | found = true; |
2020 | |
2021 | /* Now select another stripe to mark as error. */ |
2022 | failb = rbio->real_stripes - (mirror_num - 1); |
2023 | if (failb <= faila) |
2024 | failb--; |
2025 | |
2026 | /* Set the extra bit in error bitmap. */ |
2027 | if (failb >= 0) |
2028 | set_bit(nr: failb * rbio->stripe_nsectors + sector_nr, |
2029 | addr: rbio->error_bitmap); |
2030 | } |
2031 | |
2032 | /* We should found at least one vertical stripe with error.*/ |
2033 | ASSERT(found); |
2034 | } |
2035 | |
2036 | /* |
2037 | * the main entry point for reads from the higher layers. This |
2038 | * is really only called when the normal read path had a failure, |
2039 | * so we assume the bio they send down corresponds to a failed part |
2040 | * of the drive. |
2041 | */ |
2042 | void raid56_parity_recover(struct bio *bio, struct btrfs_io_context *bioc, |
2043 | int mirror_num) |
2044 | { |
2045 | struct btrfs_fs_info *fs_info = bioc->fs_info; |
2046 | struct btrfs_raid_bio *rbio; |
2047 | |
2048 | rbio = alloc_rbio(fs_info, bioc); |
2049 | if (IS_ERR(ptr: rbio)) { |
2050 | bio->bi_status = errno_to_blk_status(errno: PTR_ERR(ptr: rbio)); |
2051 | bio_endio(bio); |
2052 | return; |
2053 | } |
2054 | |
2055 | rbio->operation = BTRFS_RBIO_READ_REBUILD; |
2056 | rbio_add_bio(rbio, orig_bio: bio); |
2057 | |
2058 | set_rbio_range_error(rbio, bio); |
2059 | |
2060 | /* |
2061 | * Loop retry: |
2062 | * for 'mirror == 2', reconstruct from all other stripes. |
2063 | * for 'mirror_num > 2', select a stripe to fail on every retry. |
2064 | */ |
2065 | if (mirror_num > 2) |
2066 | set_rbio_raid6_extra_error(rbio, mirror_num); |
2067 | |
2068 | start_async_work(rbio, work_func: recover_rbio_work); |
2069 | } |
2070 | |
2071 | static void fill_data_csums(struct btrfs_raid_bio *rbio) |
2072 | { |
2073 | struct btrfs_fs_info *fs_info = rbio->bioc->fs_info; |
2074 | struct btrfs_root *csum_root = btrfs_csum_root(fs_info, |
2075 | bytenr: rbio->bioc->full_stripe_logical); |
2076 | const u64 start = rbio->bioc->full_stripe_logical; |
2077 | const u32 len = (rbio->nr_data * rbio->stripe_nsectors) << |
2078 | fs_info->sectorsize_bits; |
2079 | int ret; |
2080 | |
2081 | /* The rbio should not have its csum buffer initialized. */ |
2082 | ASSERT(!rbio->csum_buf && !rbio->csum_bitmap); |
2083 | |
2084 | /* |
2085 | * Skip the csum search if: |
2086 | * |
2087 | * - The rbio doesn't belong to data block groups |
2088 | * Then we are doing IO for tree blocks, no need to search csums. |
2089 | * |
2090 | * - The rbio belongs to mixed block groups |
2091 | * This is to avoid deadlock, as we're already holding the full |
2092 | * stripe lock, if we trigger a metadata read, and it needs to do |
2093 | * raid56 recovery, we will deadlock. |
2094 | */ |
2095 | if (!(rbio->bioc->map_type & BTRFS_BLOCK_GROUP_DATA) || |
2096 | rbio->bioc->map_type & BTRFS_BLOCK_GROUP_METADATA) |
2097 | return; |
2098 | |
2099 | rbio->csum_buf = kzalloc(size: rbio->nr_data * rbio->stripe_nsectors * |
2100 | fs_info->csum_size, GFP_NOFS); |
2101 | rbio->csum_bitmap = bitmap_zalloc(nbits: rbio->nr_data * rbio->stripe_nsectors, |
2102 | GFP_NOFS); |
2103 | if (!rbio->csum_buf || !rbio->csum_bitmap) { |
2104 | ret = -ENOMEM; |
2105 | goto error; |
2106 | } |
2107 | |
2108 | ret = btrfs_lookup_csums_bitmap(root: csum_root, NULL, start, end: start + len - 1, |
2109 | csum_buf: rbio->csum_buf, csum_bitmap: rbio->csum_bitmap); |
2110 | if (ret < 0) |
2111 | goto error; |
2112 | if (bitmap_empty(src: rbio->csum_bitmap, nbits: len >> fs_info->sectorsize_bits)) |
2113 | goto no_csum; |
2114 | return; |
2115 | |
2116 | error: |
2117 | /* |
2118 | * We failed to allocate memory or grab the csum, but it's not fatal, |
2119 | * we can still continue. But better to warn users that RMW is no |
2120 | * longer safe for this particular sub-stripe write. |
2121 | */ |
2122 | btrfs_warn_rl(fs_info, |
2123 | "sub-stripe write for full stripe %llu is not safe, failed to get csum: %d" , |
2124 | rbio->bioc->full_stripe_logical, ret); |
2125 | no_csum: |
2126 | kfree(objp: rbio->csum_buf); |
2127 | bitmap_free(bitmap: rbio->csum_bitmap); |
2128 | rbio->csum_buf = NULL; |
2129 | rbio->csum_bitmap = NULL; |
2130 | } |
2131 | |
2132 | static int rmw_read_wait_recover(struct btrfs_raid_bio *rbio) |
2133 | { |
2134 | struct bio_list bio_list = BIO_EMPTY_LIST; |
2135 | int total_sector_nr; |
2136 | int ret = 0; |
2137 | |
2138 | /* |
2139 | * Fill the data csums we need for data verification. We need to fill |
2140 | * the csum_bitmap/csum_buf first, as our endio function will try to |
2141 | * verify the data sectors. |
2142 | */ |
2143 | fill_data_csums(rbio); |
2144 | |
2145 | /* |
2146 | * Build a list of bios to read all sectors (including data and P/Q). |
2147 | * |
2148 | * This behavior is to compensate the later csum verification and recovery. |
2149 | */ |
2150 | for (total_sector_nr = 0; total_sector_nr < rbio->nr_sectors; |
2151 | total_sector_nr++) { |
2152 | struct sector_ptr *sector; |
2153 | int stripe = total_sector_nr / rbio->stripe_nsectors; |
2154 | int sectornr = total_sector_nr % rbio->stripe_nsectors; |
2155 | |
2156 | sector = rbio_stripe_sector(rbio, stripe_nr: stripe, sector_nr: sectornr); |
2157 | ret = rbio_add_io_sector(rbio, bio_list: &bio_list, sector, |
2158 | stripe_nr: stripe, sector_nr: sectornr, op: REQ_OP_READ); |
2159 | if (ret) { |
2160 | bio_list_put(bio_list: &bio_list); |
2161 | return ret; |
2162 | } |
2163 | } |
2164 | |
2165 | /* |
2166 | * We may or may not have any corrupted sectors (including missing dev |
2167 | * and csum mismatch), just let recover_sectors() to handle them all. |
2168 | */ |
2169 | submit_read_wait_bio_list(rbio, bio_list: &bio_list); |
2170 | return recover_sectors(rbio); |
2171 | } |
2172 | |
2173 | static void raid_wait_write_end_io(struct bio *bio) |
2174 | { |
2175 | struct btrfs_raid_bio *rbio = bio->bi_private; |
2176 | blk_status_t err = bio->bi_status; |
2177 | |
2178 | if (err) |
2179 | rbio_update_error_bitmap(rbio, bio); |
2180 | bio_put(bio); |
2181 | if (atomic_dec_and_test(v: &rbio->stripes_pending)) |
2182 | wake_up(&rbio->io_wait); |
2183 | } |
2184 | |
2185 | static void submit_write_bios(struct btrfs_raid_bio *rbio, |
2186 | struct bio_list *bio_list) |
2187 | { |
2188 | struct bio *bio; |
2189 | |
2190 | atomic_set(v: &rbio->stripes_pending, i: bio_list_size(bl: bio_list)); |
2191 | while ((bio = bio_list_pop(bl: bio_list))) { |
2192 | bio->bi_end_io = raid_wait_write_end_io; |
2193 | |
2194 | if (trace_raid56_write_enabled()) { |
2195 | struct raid56_bio_trace_info trace_info = { 0 }; |
2196 | |
2197 | bio_get_trace_info(rbio, bio, trace_info: &trace_info); |
2198 | trace_raid56_write(rbio, bio, trace_info: &trace_info); |
2199 | } |
2200 | submit_bio(bio); |
2201 | } |
2202 | } |
2203 | |
2204 | /* |
2205 | * To determine if we need to read any sector from the disk. |
2206 | * Should only be utilized in RMW path, to skip cached rbio. |
2207 | */ |
2208 | static bool need_read_stripe_sectors(struct btrfs_raid_bio *rbio) |
2209 | { |
2210 | int i; |
2211 | |
2212 | for (i = 0; i < rbio->nr_data * rbio->stripe_nsectors; i++) { |
2213 | struct sector_ptr *sector = &rbio->stripe_sectors[i]; |
2214 | |
2215 | /* |
2216 | * We have a sector which doesn't have page nor uptodate, |
2217 | * thus this rbio can not be cached one, as cached one must |
2218 | * have all its data sectors present and uptodate. |
2219 | */ |
2220 | if (!sector->page || !sector->uptodate) |
2221 | return true; |
2222 | } |
2223 | return false; |
2224 | } |
2225 | |
2226 | static void rmw_rbio(struct btrfs_raid_bio *rbio) |
2227 | { |
2228 | struct bio_list bio_list; |
2229 | int sectornr; |
2230 | int ret = 0; |
2231 | |
2232 | /* |
2233 | * Allocate the pages for parity first, as P/Q pages will always be |
2234 | * needed for both full-stripe and sub-stripe writes. |
2235 | */ |
2236 | ret = alloc_rbio_parity_pages(rbio); |
2237 | if (ret < 0) |
2238 | goto out; |
2239 | |
2240 | /* |
2241 | * Either full stripe write, or we have every data sector already |
2242 | * cached, can go to write path immediately. |
2243 | */ |
2244 | if (!rbio_is_full(rbio) && need_read_stripe_sectors(rbio)) { |
2245 | /* |
2246 | * Now we're doing sub-stripe write, also need all data stripes |
2247 | * to do the full RMW. |
2248 | */ |
2249 | ret = alloc_rbio_data_pages(rbio); |
2250 | if (ret < 0) |
2251 | goto out; |
2252 | |
2253 | index_rbio_pages(rbio); |
2254 | |
2255 | ret = rmw_read_wait_recover(rbio); |
2256 | if (ret < 0) |
2257 | goto out; |
2258 | } |
2259 | |
2260 | /* |
2261 | * At this stage we're not allowed to add any new bios to the |
2262 | * bio list any more, anyone else that wants to change this stripe |
2263 | * needs to do their own rmw. |
2264 | */ |
2265 | spin_lock(lock: &rbio->bio_list_lock); |
2266 | set_bit(RBIO_RMW_LOCKED_BIT, addr: &rbio->flags); |
2267 | spin_unlock(lock: &rbio->bio_list_lock); |
2268 | |
2269 | bitmap_clear(map: rbio->error_bitmap, start: 0, nbits: rbio->nr_sectors); |
2270 | |
2271 | index_rbio_pages(rbio); |
2272 | |
2273 | /* |
2274 | * We don't cache full rbios because we're assuming |
2275 | * the higher layers are unlikely to use this area of |
2276 | * the disk again soon. If they do use it again, |
2277 | * hopefully they will send another full bio. |
2278 | */ |
2279 | if (!rbio_is_full(rbio)) |
2280 | cache_rbio_pages(rbio); |
2281 | else |
2282 | clear_bit(RBIO_CACHE_READY_BIT, addr: &rbio->flags); |
2283 | |
2284 | for (sectornr = 0; sectornr < rbio->stripe_nsectors; sectornr++) |
2285 | generate_pq_vertical(rbio, sectornr); |
2286 | |
2287 | bio_list_init(bl: &bio_list); |
2288 | ret = rmw_assemble_write_bios(rbio, bio_list: &bio_list); |
2289 | if (ret < 0) |
2290 | goto out; |
2291 | |
2292 | /* We should have at least one bio assembled. */ |
2293 | ASSERT(bio_list_size(&bio_list)); |
2294 | submit_write_bios(rbio, bio_list: &bio_list); |
2295 | wait_event(rbio->io_wait, atomic_read(&rbio->stripes_pending) == 0); |
2296 | |
2297 | /* We may have more errors than our tolerance during the read. */ |
2298 | for (sectornr = 0; sectornr < rbio->stripe_nsectors; sectornr++) { |
2299 | int found_errors; |
2300 | |
2301 | found_errors = get_rbio_veritical_errors(rbio, sector_nr: sectornr, NULL, NULL); |
2302 | if (found_errors > rbio->bioc->max_errors) { |
2303 | ret = -EIO; |
2304 | break; |
2305 | } |
2306 | } |
2307 | out: |
2308 | rbio_orig_end_io(rbio, err: errno_to_blk_status(errno: ret)); |
2309 | } |
2310 | |
2311 | static void rmw_rbio_work(struct work_struct *work) |
2312 | { |
2313 | struct btrfs_raid_bio *rbio; |
2314 | |
2315 | rbio = container_of(work, struct btrfs_raid_bio, work); |
2316 | if (lock_stripe_add(rbio) == 0) |
2317 | rmw_rbio(rbio); |
2318 | } |
2319 | |
2320 | static void rmw_rbio_work_locked(struct work_struct *work) |
2321 | { |
2322 | rmw_rbio(container_of(work, struct btrfs_raid_bio, work)); |
2323 | } |
2324 | |
2325 | /* |
2326 | * The following code is used to scrub/replace the parity stripe |
2327 | * |
2328 | * Caller must have already increased bio_counter for getting @bioc. |
2329 | * |
2330 | * Note: We need make sure all the pages that add into the scrub/replace |
2331 | * raid bio are correct and not be changed during the scrub/replace. That |
2332 | * is those pages just hold metadata or file data with checksum. |
2333 | */ |
2334 | |
2335 | struct btrfs_raid_bio *raid56_parity_alloc_scrub_rbio(struct bio *bio, |
2336 | struct btrfs_io_context *bioc, |
2337 | struct btrfs_device *scrub_dev, |
2338 | unsigned long *dbitmap, int stripe_nsectors) |
2339 | { |
2340 | struct btrfs_fs_info *fs_info = bioc->fs_info; |
2341 | struct btrfs_raid_bio *rbio; |
2342 | int i; |
2343 | |
2344 | rbio = alloc_rbio(fs_info, bioc); |
2345 | if (IS_ERR(ptr: rbio)) |
2346 | return NULL; |
2347 | bio_list_add(bl: &rbio->bio_list, bio); |
2348 | /* |
2349 | * This is a special bio which is used to hold the completion handler |
2350 | * and make the scrub rbio is similar to the other types |
2351 | */ |
2352 | ASSERT(!bio->bi_iter.bi_size); |
2353 | rbio->operation = BTRFS_RBIO_PARITY_SCRUB; |
2354 | |
2355 | /* |
2356 | * After mapping bioc with BTRFS_MAP_WRITE, parities have been sorted |
2357 | * to the end position, so this search can start from the first parity |
2358 | * stripe. |
2359 | */ |
2360 | for (i = rbio->nr_data; i < rbio->real_stripes; i++) { |
2361 | if (bioc->stripes[i].dev == scrub_dev) { |
2362 | rbio->scrubp = i; |
2363 | break; |
2364 | } |
2365 | } |
2366 | ASSERT(i < rbio->real_stripes); |
2367 | |
2368 | bitmap_copy(dst: &rbio->dbitmap, src: dbitmap, nbits: stripe_nsectors); |
2369 | return rbio; |
2370 | } |
2371 | |
2372 | /* |
2373 | * We just scrub the parity that we have correct data on the same horizontal, |
2374 | * so we needn't allocate all pages for all the stripes. |
2375 | */ |
2376 | static int alloc_rbio_essential_pages(struct btrfs_raid_bio *rbio) |
2377 | { |
2378 | const u32 sectorsize = rbio->bioc->fs_info->sectorsize; |
2379 | int total_sector_nr; |
2380 | |
2381 | for (total_sector_nr = 0; total_sector_nr < rbio->nr_sectors; |
2382 | total_sector_nr++) { |
2383 | struct page *page; |
2384 | int sectornr = total_sector_nr % rbio->stripe_nsectors; |
2385 | int index = (total_sector_nr * sectorsize) >> PAGE_SHIFT; |
2386 | |
2387 | if (!test_bit(sectornr, &rbio->dbitmap)) |
2388 | continue; |
2389 | if (rbio->stripe_pages[index]) |
2390 | continue; |
2391 | page = alloc_page(GFP_NOFS); |
2392 | if (!page) |
2393 | return -ENOMEM; |
2394 | rbio->stripe_pages[index] = page; |
2395 | } |
2396 | index_stripe_sectors(rbio); |
2397 | return 0; |
2398 | } |
2399 | |
2400 | static int finish_parity_scrub(struct btrfs_raid_bio *rbio) |
2401 | { |
2402 | struct btrfs_io_context *bioc = rbio->bioc; |
2403 | const u32 sectorsize = bioc->fs_info->sectorsize; |
2404 | void **pointers = rbio->finish_pointers; |
2405 | unsigned long *pbitmap = &rbio->finish_pbitmap; |
2406 | int nr_data = rbio->nr_data; |
2407 | int stripe; |
2408 | int sectornr; |
2409 | bool has_qstripe; |
2410 | struct sector_ptr p_sector = { 0 }; |
2411 | struct sector_ptr q_sector = { 0 }; |
2412 | struct bio_list bio_list; |
2413 | int is_replace = 0; |
2414 | int ret; |
2415 | |
2416 | bio_list_init(bl: &bio_list); |
2417 | |
2418 | if (rbio->real_stripes - rbio->nr_data == 1) |
2419 | has_qstripe = false; |
2420 | else if (rbio->real_stripes - rbio->nr_data == 2) |
2421 | has_qstripe = true; |
2422 | else |
2423 | BUG(); |
2424 | |
2425 | /* |
2426 | * Replace is running and our P/Q stripe is being replaced, then we |
2427 | * need to duplicate the final write to replace target. |
2428 | */ |
2429 | if (bioc->replace_nr_stripes && bioc->replace_stripe_src == rbio->scrubp) { |
2430 | is_replace = 1; |
2431 | bitmap_copy(dst: pbitmap, src: &rbio->dbitmap, nbits: rbio->stripe_nsectors); |
2432 | } |
2433 | |
2434 | /* |
2435 | * Because the higher layers(scrubber) are unlikely to |
2436 | * use this area of the disk again soon, so don't cache |
2437 | * it. |
2438 | */ |
2439 | clear_bit(RBIO_CACHE_READY_BIT, addr: &rbio->flags); |
2440 | |
2441 | p_sector.page = alloc_page(GFP_NOFS); |
2442 | if (!p_sector.page) |
2443 | return -ENOMEM; |
2444 | p_sector.pgoff = 0; |
2445 | p_sector.uptodate = 1; |
2446 | |
2447 | if (has_qstripe) { |
2448 | /* RAID6, allocate and map temp space for the Q stripe */ |
2449 | q_sector.page = alloc_page(GFP_NOFS); |
2450 | if (!q_sector.page) { |
2451 | __free_page(p_sector.page); |
2452 | p_sector.page = NULL; |
2453 | return -ENOMEM; |
2454 | } |
2455 | q_sector.pgoff = 0; |
2456 | q_sector.uptodate = 1; |
2457 | pointers[rbio->real_stripes - 1] = kmap_local_page(page: q_sector.page); |
2458 | } |
2459 | |
2460 | bitmap_clear(map: rbio->error_bitmap, start: 0, nbits: rbio->nr_sectors); |
2461 | |
2462 | /* Map the parity stripe just once */ |
2463 | pointers[nr_data] = kmap_local_page(page: p_sector.page); |
2464 | |
2465 | for_each_set_bit(sectornr, &rbio->dbitmap, rbio->stripe_nsectors) { |
2466 | struct sector_ptr *sector; |
2467 | void *parity; |
2468 | |
2469 | /* first collect one page from each data stripe */ |
2470 | for (stripe = 0; stripe < nr_data; stripe++) { |
2471 | sector = sector_in_rbio(rbio, stripe_nr: stripe, sector_nr: sectornr, bio_list_only: 0); |
2472 | pointers[stripe] = kmap_local_page(page: sector->page) + |
2473 | sector->pgoff; |
2474 | } |
2475 | |
2476 | if (has_qstripe) { |
2477 | /* RAID6, call the library function to fill in our P/Q */ |
2478 | raid6_call.gen_syndrome(rbio->real_stripes, sectorsize, |
2479 | pointers); |
2480 | } else { |
2481 | /* raid5 */ |
2482 | memcpy(pointers[nr_data], pointers[0], sectorsize); |
2483 | run_xor(pages: pointers + 1, src_cnt: nr_data - 1, len: sectorsize); |
2484 | } |
2485 | |
2486 | /* Check scrubbing parity and repair it */ |
2487 | sector = rbio_stripe_sector(rbio, stripe_nr: rbio->scrubp, sector_nr: sectornr); |
2488 | parity = kmap_local_page(page: sector->page) + sector->pgoff; |
2489 | if (memcmp(p: parity, q: pointers[rbio->scrubp], size: sectorsize) != 0) |
2490 | memcpy(parity, pointers[rbio->scrubp], sectorsize); |
2491 | else |
2492 | /* Parity is right, needn't writeback */ |
2493 | bitmap_clear(map: &rbio->dbitmap, start: sectornr, nbits: 1); |
2494 | kunmap_local(parity); |
2495 | |
2496 | for (stripe = nr_data - 1; stripe >= 0; stripe--) |
2497 | kunmap_local(pointers[stripe]); |
2498 | } |
2499 | |
2500 | kunmap_local(pointers[nr_data]); |
2501 | __free_page(p_sector.page); |
2502 | p_sector.page = NULL; |
2503 | if (q_sector.page) { |
2504 | kunmap_local(pointers[rbio->real_stripes - 1]); |
2505 | __free_page(q_sector.page); |
2506 | q_sector.page = NULL; |
2507 | } |
2508 | |
2509 | /* |
2510 | * time to start writing. Make bios for everything from the |
2511 | * higher layers (the bio_list in our rbio) and our p/q. Ignore |
2512 | * everything else. |
2513 | */ |
2514 | for_each_set_bit(sectornr, &rbio->dbitmap, rbio->stripe_nsectors) { |
2515 | struct sector_ptr *sector; |
2516 | |
2517 | sector = rbio_stripe_sector(rbio, stripe_nr: rbio->scrubp, sector_nr: sectornr); |
2518 | ret = rbio_add_io_sector(rbio, bio_list: &bio_list, sector, stripe_nr: rbio->scrubp, |
2519 | sector_nr: sectornr, op: REQ_OP_WRITE); |
2520 | if (ret) |
2521 | goto cleanup; |
2522 | } |
2523 | |
2524 | if (!is_replace) |
2525 | goto submit_write; |
2526 | |
2527 | /* |
2528 | * Replace is running and our parity stripe needs to be duplicated to |
2529 | * the target device. Check we have a valid source stripe number. |
2530 | */ |
2531 | ASSERT(rbio->bioc->replace_stripe_src >= 0); |
2532 | for_each_set_bit(sectornr, pbitmap, rbio->stripe_nsectors) { |
2533 | struct sector_ptr *sector; |
2534 | |
2535 | sector = rbio_stripe_sector(rbio, stripe_nr: rbio->scrubp, sector_nr: sectornr); |
2536 | ret = rbio_add_io_sector(rbio, bio_list: &bio_list, sector, |
2537 | stripe_nr: rbio->real_stripes, |
2538 | sector_nr: sectornr, op: REQ_OP_WRITE); |
2539 | if (ret) |
2540 | goto cleanup; |
2541 | } |
2542 | |
2543 | submit_write: |
2544 | submit_write_bios(rbio, bio_list: &bio_list); |
2545 | return 0; |
2546 | |
2547 | cleanup: |
2548 | bio_list_put(bio_list: &bio_list); |
2549 | return ret; |
2550 | } |
2551 | |
2552 | static inline int is_data_stripe(struct btrfs_raid_bio *rbio, int stripe) |
2553 | { |
2554 | if (stripe >= 0 && stripe < rbio->nr_data) |
2555 | return 1; |
2556 | return 0; |
2557 | } |
2558 | |
2559 | static int recover_scrub_rbio(struct btrfs_raid_bio *rbio) |
2560 | { |
2561 | void **pointers = NULL; |
2562 | void **unmap_array = NULL; |
2563 | int sector_nr; |
2564 | int ret = 0; |
2565 | |
2566 | /* |
2567 | * @pointers array stores the pointer for each sector. |
2568 | * |
2569 | * @unmap_array stores copy of pointers that does not get reordered |
2570 | * during reconstruction so that kunmap_local works. |
2571 | */ |
2572 | pointers = kcalloc(n: rbio->real_stripes, size: sizeof(void *), GFP_NOFS); |
2573 | unmap_array = kcalloc(n: rbio->real_stripes, size: sizeof(void *), GFP_NOFS); |
2574 | if (!pointers || !unmap_array) { |
2575 | ret = -ENOMEM; |
2576 | goto out; |
2577 | } |
2578 | |
2579 | for (sector_nr = 0; sector_nr < rbio->stripe_nsectors; sector_nr++) { |
2580 | int dfail = 0, failp = -1; |
2581 | int faila; |
2582 | int failb; |
2583 | int found_errors; |
2584 | |
2585 | found_errors = get_rbio_veritical_errors(rbio, sector_nr, |
2586 | faila: &faila, failb: &failb); |
2587 | if (found_errors > rbio->bioc->max_errors) { |
2588 | ret = -EIO; |
2589 | goto out; |
2590 | } |
2591 | if (found_errors == 0) |
2592 | continue; |
2593 | |
2594 | /* We should have at least one error here. */ |
2595 | ASSERT(faila >= 0 || failb >= 0); |
2596 | |
2597 | if (is_data_stripe(rbio, stripe: faila)) |
2598 | dfail++; |
2599 | else if (is_parity_stripe(faila)) |
2600 | failp = faila; |
2601 | |
2602 | if (is_data_stripe(rbio, stripe: failb)) |
2603 | dfail++; |
2604 | else if (is_parity_stripe(failb)) |
2605 | failp = failb; |
2606 | /* |
2607 | * Because we can not use a scrubbing parity to repair the |
2608 | * data, so the capability of the repair is declined. (In the |
2609 | * case of RAID5, we can not repair anything.) |
2610 | */ |
2611 | if (dfail > rbio->bioc->max_errors - 1) { |
2612 | ret = -EIO; |
2613 | goto out; |
2614 | } |
2615 | /* |
2616 | * If all data is good, only parity is correctly, just repair |
2617 | * the parity, no need to recover data stripes. |
2618 | */ |
2619 | if (dfail == 0) |
2620 | continue; |
2621 | |
2622 | /* |
2623 | * Here means we got one corrupted data stripe and one |
2624 | * corrupted parity on RAID6, if the corrupted parity is |
2625 | * scrubbing parity, luckily, use the other one to repair the |
2626 | * data, or we can not repair the data stripe. |
2627 | */ |
2628 | if (failp != rbio->scrubp) { |
2629 | ret = -EIO; |
2630 | goto out; |
2631 | } |
2632 | |
2633 | ret = recover_vertical(rbio, sector_nr, pointers, unmap_array); |
2634 | if (ret < 0) |
2635 | goto out; |
2636 | } |
2637 | out: |
2638 | kfree(objp: pointers); |
2639 | kfree(objp: unmap_array); |
2640 | return ret; |
2641 | } |
2642 | |
2643 | static int scrub_assemble_read_bios(struct btrfs_raid_bio *rbio) |
2644 | { |
2645 | struct bio_list bio_list = BIO_EMPTY_LIST; |
2646 | int total_sector_nr; |
2647 | int ret = 0; |
2648 | |
2649 | /* Build a list of bios to read all the missing parts. */ |
2650 | for (total_sector_nr = 0; total_sector_nr < rbio->nr_sectors; |
2651 | total_sector_nr++) { |
2652 | int sectornr = total_sector_nr % rbio->stripe_nsectors; |
2653 | int stripe = total_sector_nr / rbio->stripe_nsectors; |
2654 | struct sector_ptr *sector; |
2655 | |
2656 | /* No data in the vertical stripe, no need to read. */ |
2657 | if (!test_bit(sectornr, &rbio->dbitmap)) |
2658 | continue; |
2659 | |
2660 | /* |
2661 | * We want to find all the sectors missing from the rbio and |
2662 | * read them from the disk. If sector_in_rbio() finds a sector |
2663 | * in the bio list we don't need to read it off the stripe. |
2664 | */ |
2665 | sector = sector_in_rbio(rbio, stripe_nr: stripe, sector_nr: sectornr, bio_list_only: 1); |
2666 | if (sector) |
2667 | continue; |
2668 | |
2669 | sector = rbio_stripe_sector(rbio, stripe_nr: stripe, sector_nr: sectornr); |
2670 | /* |
2671 | * The bio cache may have handed us an uptodate sector. If so, |
2672 | * use it. |
2673 | */ |
2674 | if (sector->uptodate) |
2675 | continue; |
2676 | |
2677 | ret = rbio_add_io_sector(rbio, bio_list: &bio_list, sector, stripe_nr: stripe, |
2678 | sector_nr: sectornr, op: REQ_OP_READ); |
2679 | if (ret) { |
2680 | bio_list_put(bio_list: &bio_list); |
2681 | return ret; |
2682 | } |
2683 | } |
2684 | |
2685 | submit_read_wait_bio_list(rbio, bio_list: &bio_list); |
2686 | return 0; |
2687 | } |
2688 | |
2689 | static void scrub_rbio(struct btrfs_raid_bio *rbio) |
2690 | { |
2691 | int sector_nr; |
2692 | int ret; |
2693 | |
2694 | ret = alloc_rbio_essential_pages(rbio); |
2695 | if (ret) |
2696 | goto out; |
2697 | |
2698 | bitmap_clear(map: rbio->error_bitmap, start: 0, nbits: rbio->nr_sectors); |
2699 | |
2700 | ret = scrub_assemble_read_bios(rbio); |
2701 | if (ret < 0) |
2702 | goto out; |
2703 | |
2704 | /* We may have some failures, recover the failed sectors first. */ |
2705 | ret = recover_scrub_rbio(rbio); |
2706 | if (ret < 0) |
2707 | goto out; |
2708 | |
2709 | /* |
2710 | * We have every sector properly prepared. Can finish the scrub |
2711 | * and writeback the good content. |
2712 | */ |
2713 | ret = finish_parity_scrub(rbio); |
2714 | wait_event(rbio->io_wait, atomic_read(&rbio->stripes_pending) == 0); |
2715 | for (sector_nr = 0; sector_nr < rbio->stripe_nsectors; sector_nr++) { |
2716 | int found_errors; |
2717 | |
2718 | found_errors = get_rbio_veritical_errors(rbio, sector_nr, NULL, NULL); |
2719 | if (found_errors > rbio->bioc->max_errors) { |
2720 | ret = -EIO; |
2721 | break; |
2722 | } |
2723 | } |
2724 | out: |
2725 | rbio_orig_end_io(rbio, err: errno_to_blk_status(errno: ret)); |
2726 | } |
2727 | |
2728 | static void scrub_rbio_work_locked(struct work_struct *work) |
2729 | { |
2730 | scrub_rbio(container_of(work, struct btrfs_raid_bio, work)); |
2731 | } |
2732 | |
2733 | void raid56_parity_submit_scrub_rbio(struct btrfs_raid_bio *rbio) |
2734 | { |
2735 | if (!lock_stripe_add(rbio)) |
2736 | start_async_work(rbio, work_func: scrub_rbio_work_locked); |
2737 | } |
2738 | |
2739 | /* |
2740 | * This is for scrub call sites where we already have correct data contents. |
2741 | * This allows us to avoid reading data stripes again. |
2742 | * |
2743 | * Unfortunately here we have to do page copy, other than reusing the pages. |
2744 | * This is due to the fact rbio has its own page management for its cache. |
2745 | */ |
2746 | void raid56_parity_cache_data_pages(struct btrfs_raid_bio *rbio, |
2747 | struct page **data_pages, u64 data_logical) |
2748 | { |
2749 | const u64 offset_in_full_stripe = data_logical - |
2750 | rbio->bioc->full_stripe_logical; |
2751 | const int page_index = offset_in_full_stripe >> PAGE_SHIFT; |
2752 | const u32 sectorsize = rbio->bioc->fs_info->sectorsize; |
2753 | const u32 sectors_per_page = PAGE_SIZE / sectorsize; |
2754 | int ret; |
2755 | |
2756 | /* |
2757 | * If we hit ENOMEM temporarily, but later at |
2758 | * raid56_parity_submit_scrub_rbio() time it succeeded, we just do |
2759 | * the extra read, not a big deal. |
2760 | * |
2761 | * If we hit ENOMEM later at raid56_parity_submit_scrub_rbio() time, |
2762 | * the bio would got proper error number set. |
2763 | */ |
2764 | ret = alloc_rbio_data_pages(rbio); |
2765 | if (ret < 0) |
2766 | return; |
2767 | |
2768 | /* data_logical must be at stripe boundary and inside the full stripe. */ |
2769 | ASSERT(IS_ALIGNED(offset_in_full_stripe, BTRFS_STRIPE_LEN)); |
2770 | ASSERT(offset_in_full_stripe < (rbio->nr_data << BTRFS_STRIPE_LEN_SHIFT)); |
2771 | |
2772 | for (int page_nr = 0; page_nr < (BTRFS_STRIPE_LEN >> PAGE_SHIFT); page_nr++) { |
2773 | struct page *dst = rbio->stripe_pages[page_nr + page_index]; |
2774 | struct page *src = data_pages[page_nr]; |
2775 | |
2776 | memcpy_page(dst_page: dst, dst_off: 0, src_page: src, src_off: 0, PAGE_SIZE); |
2777 | for (int sector_nr = sectors_per_page * page_index; |
2778 | sector_nr < sectors_per_page * (page_index + 1); |
2779 | sector_nr++) |
2780 | rbio->stripe_sectors[sector_nr].uptodate = true; |
2781 | } |
2782 | } |
2783 | |