Loading...
1// SPDX-License-Identifier: GPL-2.0
2/*
3 * Copyright (C) 2007 Oracle. All rights reserved.
4 * Copyright (C) 2022 Christoph Hellwig.
5 */
6
7#include <linux/bio.h>
8#include "bio.h"
9#include "ctree.h"
10#include "volumes.h"
11#include "raid56.h"
12#include "async-thread.h"
13#include "dev-replace.h"
14#include "zoned.h"
15#include "file-item.h"
16#include "raid-stripe-tree.h"
17
18static struct bio_set btrfs_bioset;
19static struct bio_set btrfs_clone_bioset;
20static struct bio_set btrfs_repair_bioset;
21static mempool_t btrfs_failed_bio_pool;
22
23struct btrfs_failed_bio {
24 struct btrfs_bio *bbio;
25 int num_copies;
26 atomic_t repair_count;
27};
28
29/* Is this a data path I/O that needs storage layer checksum and repair? */
30static inline bool is_data_bbio(struct btrfs_bio *bbio)
31{
32 return bbio->inode && is_data_inode(bbio->inode);
33}
34
35static bool bbio_has_ordered_extent(struct btrfs_bio *bbio)
36{
37 return is_data_bbio(bbio) && btrfs_op(&bbio->bio) == BTRFS_MAP_WRITE;
38}
39
40/*
41 * Initialize a btrfs_bio structure. This skips the embedded bio itself as it
42 * is already initialized by the block layer.
43 */
44void btrfs_bio_init(struct btrfs_bio *bbio, struct btrfs_fs_info *fs_info,
45 btrfs_bio_end_io_t end_io, void *private)
46{
47 memset(bbio, 0, offsetof(struct btrfs_bio, bio));
48 bbio->fs_info = fs_info;
49 bbio->end_io = end_io;
50 bbio->private = private;
51 atomic_set(&bbio->pending_ios, 1);
52 WRITE_ONCE(bbio->status, BLK_STS_OK);
53}
54
55/*
56 * Allocate a btrfs_bio structure. The btrfs_bio is the main I/O container for
57 * btrfs, and is used for all I/O submitted through btrfs_submit_bbio().
58 *
59 * Just like the underlying bio_alloc_bioset it will not fail as it is backed by
60 * a mempool.
61 */
62struct btrfs_bio *btrfs_bio_alloc(unsigned int nr_vecs, blk_opf_t opf,
63 struct btrfs_fs_info *fs_info,
64 btrfs_bio_end_io_t end_io, void *private)
65{
66 struct btrfs_bio *bbio;
67 struct bio *bio;
68
69 bio = bio_alloc_bioset(NULL, nr_vecs, opf, GFP_NOFS, &btrfs_bioset);
70 bbio = btrfs_bio(bio);
71 btrfs_bio_init(bbio, fs_info, end_io, private);
72 return bbio;
73}
74
75static struct btrfs_bio *btrfs_split_bio(struct btrfs_fs_info *fs_info,
76 struct btrfs_bio *orig_bbio,
77 u64 map_length)
78{
79 struct btrfs_bio *bbio;
80 struct bio *bio;
81
82 bio = bio_split(&orig_bbio->bio, map_length >> SECTOR_SHIFT, GFP_NOFS,
83 &btrfs_clone_bioset);
84 if (IS_ERR(bio))
85 return ERR_CAST(bio);
86
87 bbio = btrfs_bio(bio);
88 btrfs_bio_init(bbio, fs_info, NULL, orig_bbio);
89 bbio->inode = orig_bbio->inode;
90 bbio->file_offset = orig_bbio->file_offset;
91 orig_bbio->file_offset += map_length;
92 if (bbio_has_ordered_extent(bbio)) {
93 refcount_inc(&orig_bbio->ordered->refs);
94 bbio->ordered = orig_bbio->ordered;
95 }
96 atomic_inc(&orig_bbio->pending_ios);
97 return bbio;
98}
99
100/* Free a bio that was never submitted to the underlying device. */
101static void btrfs_cleanup_bio(struct btrfs_bio *bbio)
102{
103 if (bbio_has_ordered_extent(bbio))
104 btrfs_put_ordered_extent(bbio->ordered);
105 bio_put(&bbio->bio);
106}
107
108static void __btrfs_bio_end_io(struct btrfs_bio *bbio)
109{
110 if (bbio_has_ordered_extent(bbio)) {
111 struct btrfs_ordered_extent *ordered = bbio->ordered;
112
113 bbio->end_io(bbio);
114 btrfs_put_ordered_extent(ordered);
115 } else {
116 bbio->end_io(bbio);
117 }
118}
119
120void btrfs_bio_end_io(struct btrfs_bio *bbio, blk_status_t status)
121{
122 bbio->bio.bi_status = status;
123 if (bbio->bio.bi_pool == &btrfs_clone_bioset) {
124 struct btrfs_bio *orig_bbio = bbio->private;
125
126 btrfs_cleanup_bio(bbio);
127 bbio = orig_bbio;
128 }
129
130 /*
131 * At this point, bbio always points to the original btrfs_bio. Save
132 * the first error in it.
133 */
134 if (status != BLK_STS_OK)
135 cmpxchg(&bbio->status, BLK_STS_OK, status);
136
137 if (atomic_dec_and_test(&bbio->pending_ios)) {
138 /* Load split bio's error which might be set above. */
139 if (status == BLK_STS_OK)
140 bbio->bio.bi_status = READ_ONCE(bbio->status);
141 __btrfs_bio_end_io(bbio);
142 }
143}
144
145static int next_repair_mirror(struct btrfs_failed_bio *fbio, int cur_mirror)
146{
147 if (cur_mirror == fbio->num_copies)
148 return cur_mirror + 1 - fbio->num_copies;
149 return cur_mirror + 1;
150}
151
152static int prev_repair_mirror(struct btrfs_failed_bio *fbio, int cur_mirror)
153{
154 if (cur_mirror == 1)
155 return fbio->num_copies;
156 return cur_mirror - 1;
157}
158
159static void btrfs_repair_done(struct btrfs_failed_bio *fbio)
160{
161 if (atomic_dec_and_test(&fbio->repair_count)) {
162 btrfs_bio_end_io(fbio->bbio, fbio->bbio->bio.bi_status);
163 mempool_free(fbio, &btrfs_failed_bio_pool);
164 }
165}
166
167static void btrfs_end_repair_bio(struct btrfs_bio *repair_bbio,
168 struct btrfs_device *dev)
169{
170 struct btrfs_failed_bio *fbio = repair_bbio->private;
171 struct btrfs_inode *inode = repair_bbio->inode;
172 struct btrfs_fs_info *fs_info = inode->root->fs_info;
173 struct bio_vec *bv = bio_first_bvec_all(&repair_bbio->bio);
174 int mirror = repair_bbio->mirror_num;
175
176 /*
177 * We can only trigger this for data bio, which doesn't support larger
178 * folios yet.
179 */
180 ASSERT(folio_order(page_folio(bv->bv_page)) == 0);
181
182 if (repair_bbio->bio.bi_status ||
183 !btrfs_data_csum_ok(repair_bbio, dev, 0, bv)) {
184 bio_reset(&repair_bbio->bio, NULL, REQ_OP_READ);
185 repair_bbio->bio.bi_iter = repair_bbio->saved_iter;
186
187 mirror = next_repair_mirror(fbio, mirror);
188 if (mirror == fbio->bbio->mirror_num) {
189 btrfs_debug(fs_info, "no mirror left");
190 fbio->bbio->bio.bi_status = BLK_STS_IOERR;
191 goto done;
192 }
193
194 btrfs_submit_bbio(repair_bbio, mirror);
195 return;
196 }
197
198 do {
199 mirror = prev_repair_mirror(fbio, mirror);
200 btrfs_repair_io_failure(fs_info, btrfs_ino(inode),
201 repair_bbio->file_offset, fs_info->sectorsize,
202 repair_bbio->saved_iter.bi_sector << SECTOR_SHIFT,
203 page_folio(bv->bv_page), bv->bv_offset, mirror);
204 } while (mirror != fbio->bbio->mirror_num);
205
206done:
207 btrfs_repair_done(fbio);
208 bio_put(&repair_bbio->bio);
209}
210
211/*
212 * Try to kick off a repair read to the next available mirror for a bad sector.
213 *
214 * This primarily tries to recover good data to serve the actual read request,
215 * but also tries to write the good data back to the bad mirror(s) when a
216 * read succeeded to restore the redundancy.
217 */
218static struct btrfs_failed_bio *repair_one_sector(struct btrfs_bio *failed_bbio,
219 u32 bio_offset,
220 struct bio_vec *bv,
221 struct btrfs_failed_bio *fbio)
222{
223 struct btrfs_inode *inode = failed_bbio->inode;
224 struct btrfs_fs_info *fs_info = inode->root->fs_info;
225 const u32 sectorsize = fs_info->sectorsize;
226 const u64 logical = (failed_bbio->saved_iter.bi_sector << SECTOR_SHIFT);
227 struct btrfs_bio *repair_bbio;
228 struct bio *repair_bio;
229 int num_copies;
230 int mirror;
231
232 btrfs_debug(fs_info, "repair read error: read error at %llu",
233 failed_bbio->file_offset + bio_offset);
234
235 num_copies = btrfs_num_copies(fs_info, logical, sectorsize);
236 if (num_copies == 1) {
237 btrfs_debug(fs_info, "no copy to repair from");
238 failed_bbio->bio.bi_status = BLK_STS_IOERR;
239 return fbio;
240 }
241
242 if (!fbio) {
243 fbio = mempool_alloc(&btrfs_failed_bio_pool, GFP_NOFS);
244 fbio->bbio = failed_bbio;
245 fbio->num_copies = num_copies;
246 atomic_set(&fbio->repair_count, 1);
247 }
248
249 atomic_inc(&fbio->repair_count);
250
251 repair_bio = bio_alloc_bioset(NULL, 1, REQ_OP_READ, GFP_NOFS,
252 &btrfs_repair_bioset);
253 repair_bio->bi_iter.bi_sector = failed_bbio->saved_iter.bi_sector;
254 __bio_add_page(repair_bio, bv->bv_page, bv->bv_len, bv->bv_offset);
255
256 repair_bbio = btrfs_bio(repair_bio);
257 btrfs_bio_init(repair_bbio, fs_info, NULL, fbio);
258 repair_bbio->inode = failed_bbio->inode;
259 repair_bbio->file_offset = failed_bbio->file_offset + bio_offset;
260
261 mirror = next_repair_mirror(fbio, failed_bbio->mirror_num);
262 btrfs_debug(fs_info, "submitting repair read to mirror %d", mirror);
263 btrfs_submit_bbio(repair_bbio, mirror);
264 return fbio;
265}
266
267static void btrfs_check_read_bio(struct btrfs_bio *bbio, struct btrfs_device *dev)
268{
269 struct btrfs_inode *inode = bbio->inode;
270 struct btrfs_fs_info *fs_info = inode->root->fs_info;
271 u32 sectorsize = fs_info->sectorsize;
272 struct bvec_iter *iter = &bbio->saved_iter;
273 blk_status_t status = bbio->bio.bi_status;
274 struct btrfs_failed_bio *fbio = NULL;
275 u32 offset = 0;
276
277 /* Read-repair requires the inode field to be set by the submitter. */
278 ASSERT(inode);
279
280 /*
281 * Hand off repair bios to the repair code as there is no upper level
282 * submitter for them.
283 */
284 if (bbio->bio.bi_pool == &btrfs_repair_bioset) {
285 btrfs_end_repair_bio(bbio, dev);
286 return;
287 }
288
289 /* Clear the I/O error. A failed repair will reset it. */
290 bbio->bio.bi_status = BLK_STS_OK;
291
292 while (iter->bi_size) {
293 struct bio_vec bv = bio_iter_iovec(&bbio->bio, *iter);
294
295 bv.bv_len = min(bv.bv_len, sectorsize);
296 if (status || !btrfs_data_csum_ok(bbio, dev, offset, &bv))
297 fbio = repair_one_sector(bbio, offset, &bv, fbio);
298
299 bio_advance_iter_single(&bbio->bio, iter, sectorsize);
300 offset += sectorsize;
301 }
302
303 if (bbio->csum != bbio->csum_inline)
304 kfree(bbio->csum);
305
306 if (fbio)
307 btrfs_repair_done(fbio);
308 else
309 btrfs_bio_end_io(bbio, bbio->bio.bi_status);
310}
311
312static void btrfs_log_dev_io_error(struct bio *bio, struct btrfs_device *dev)
313{
314 if (!dev || !dev->bdev)
315 return;
316 if (bio->bi_status != BLK_STS_IOERR && bio->bi_status != BLK_STS_TARGET)
317 return;
318
319 if (btrfs_op(bio) == BTRFS_MAP_WRITE)
320 btrfs_dev_stat_inc_and_print(dev, BTRFS_DEV_STAT_WRITE_ERRS);
321 else if (!(bio->bi_opf & REQ_RAHEAD))
322 btrfs_dev_stat_inc_and_print(dev, BTRFS_DEV_STAT_READ_ERRS);
323 if (bio->bi_opf & REQ_PREFLUSH)
324 btrfs_dev_stat_inc_and_print(dev, BTRFS_DEV_STAT_FLUSH_ERRS);
325}
326
327static struct workqueue_struct *btrfs_end_io_wq(struct btrfs_fs_info *fs_info,
328 struct bio *bio)
329{
330 if (bio->bi_opf & REQ_META)
331 return fs_info->endio_meta_workers;
332 return fs_info->endio_workers;
333}
334
335static void btrfs_end_bio_work(struct work_struct *work)
336{
337 struct btrfs_bio *bbio = container_of(work, struct btrfs_bio, end_io_work);
338
339 /* Metadata reads are checked and repaired by the submitter. */
340 if (is_data_bbio(bbio))
341 btrfs_check_read_bio(bbio, bbio->bio.bi_private);
342 else
343 btrfs_bio_end_io(bbio, bbio->bio.bi_status);
344}
345
346static void btrfs_simple_end_io(struct bio *bio)
347{
348 struct btrfs_bio *bbio = btrfs_bio(bio);
349 struct btrfs_device *dev = bio->bi_private;
350 struct btrfs_fs_info *fs_info = bbio->fs_info;
351
352 btrfs_bio_counter_dec(fs_info);
353
354 if (bio->bi_status)
355 btrfs_log_dev_io_error(bio, dev);
356
357 if (bio_op(bio) == REQ_OP_READ) {
358 INIT_WORK(&bbio->end_io_work, btrfs_end_bio_work);
359 queue_work(btrfs_end_io_wq(fs_info, bio), &bbio->end_io_work);
360 } else {
361 if (bio_is_zone_append(bio) && !bio->bi_status)
362 btrfs_record_physical_zoned(bbio);
363 btrfs_bio_end_io(bbio, bbio->bio.bi_status);
364 }
365}
366
367static void btrfs_raid56_end_io(struct bio *bio)
368{
369 struct btrfs_io_context *bioc = bio->bi_private;
370 struct btrfs_bio *bbio = btrfs_bio(bio);
371
372 btrfs_bio_counter_dec(bioc->fs_info);
373 bbio->mirror_num = bioc->mirror_num;
374 if (bio_op(bio) == REQ_OP_READ && is_data_bbio(bbio))
375 btrfs_check_read_bio(bbio, NULL);
376 else
377 btrfs_bio_end_io(bbio, bbio->bio.bi_status);
378
379 btrfs_put_bioc(bioc);
380}
381
382static void btrfs_orig_write_end_io(struct bio *bio)
383{
384 struct btrfs_io_stripe *stripe = bio->bi_private;
385 struct btrfs_io_context *bioc = stripe->bioc;
386 struct btrfs_bio *bbio = btrfs_bio(bio);
387
388 btrfs_bio_counter_dec(bioc->fs_info);
389
390 if (bio->bi_status) {
391 atomic_inc(&bioc->error);
392 btrfs_log_dev_io_error(bio, stripe->dev);
393 }
394
395 /*
396 * Only send an error to the higher layers if it is beyond the tolerance
397 * threshold.
398 */
399 if (atomic_read(&bioc->error) > bioc->max_errors)
400 bio->bi_status = BLK_STS_IOERR;
401 else
402 bio->bi_status = BLK_STS_OK;
403
404 if (bio_is_zone_append(bio) && !bio->bi_status)
405 stripe->physical = bio->bi_iter.bi_sector << SECTOR_SHIFT;
406
407 btrfs_bio_end_io(bbio, bbio->bio.bi_status);
408 btrfs_put_bioc(bioc);
409}
410
411static void btrfs_clone_write_end_io(struct bio *bio)
412{
413 struct btrfs_io_stripe *stripe = bio->bi_private;
414
415 if (bio->bi_status) {
416 atomic_inc(&stripe->bioc->error);
417 btrfs_log_dev_io_error(bio, stripe->dev);
418 } else if (bio_is_zone_append(bio)) {
419 stripe->physical = bio->bi_iter.bi_sector << SECTOR_SHIFT;
420 }
421
422 /* Pass on control to the original bio this one was cloned from */
423 bio_endio(stripe->bioc->orig_bio);
424 bio_put(bio);
425}
426
427static void btrfs_submit_dev_bio(struct btrfs_device *dev, struct bio *bio)
428{
429 if (!dev || !dev->bdev ||
430 test_bit(BTRFS_DEV_STATE_MISSING, &dev->dev_state) ||
431 (btrfs_op(bio) == BTRFS_MAP_WRITE &&
432 !test_bit(BTRFS_DEV_STATE_WRITEABLE, &dev->dev_state))) {
433 bio_io_error(bio);
434 return;
435 }
436
437 bio_set_dev(bio, dev->bdev);
438
439 /*
440 * For zone append writing, bi_sector must point the beginning of the
441 * zone
442 */
443 if (bio_op(bio) == REQ_OP_ZONE_APPEND) {
444 u64 physical = bio->bi_iter.bi_sector << SECTOR_SHIFT;
445 u64 zone_start = round_down(physical, dev->fs_info->zone_size);
446
447 ASSERT(btrfs_dev_is_sequential(dev, physical));
448 bio->bi_iter.bi_sector = zone_start >> SECTOR_SHIFT;
449 }
450 btrfs_debug_in_rcu(dev->fs_info,
451 "%s: rw %d 0x%x, sector=%llu, dev=%lu (%s id %llu), size=%u",
452 __func__, bio_op(bio), bio->bi_opf, bio->bi_iter.bi_sector,
453 (unsigned long)dev->bdev->bd_dev, btrfs_dev_name(dev),
454 dev->devid, bio->bi_iter.bi_size);
455
456 if (bio->bi_opf & REQ_BTRFS_CGROUP_PUNT)
457 blkcg_punt_bio_submit(bio);
458 else
459 submit_bio(bio);
460}
461
462static void btrfs_submit_mirrored_bio(struct btrfs_io_context *bioc, int dev_nr)
463{
464 struct bio *orig_bio = bioc->orig_bio, *bio;
465
466 ASSERT(bio_op(orig_bio) != REQ_OP_READ);
467
468 /* Reuse the bio embedded into the btrfs_bio for the last mirror */
469 if (dev_nr == bioc->num_stripes - 1) {
470 bio = orig_bio;
471 bio->bi_end_io = btrfs_orig_write_end_io;
472 } else {
473 bio = bio_alloc_clone(NULL, orig_bio, GFP_NOFS, &fs_bio_set);
474 bio_inc_remaining(orig_bio);
475 bio->bi_end_io = btrfs_clone_write_end_io;
476 }
477
478 bio->bi_private = &bioc->stripes[dev_nr];
479 bio->bi_iter.bi_sector = bioc->stripes[dev_nr].physical >> SECTOR_SHIFT;
480 bioc->stripes[dev_nr].bioc = bioc;
481 bioc->size = bio->bi_iter.bi_size;
482 btrfs_submit_dev_bio(bioc->stripes[dev_nr].dev, bio);
483}
484
485static void btrfs_submit_bio(struct bio *bio, struct btrfs_io_context *bioc,
486 struct btrfs_io_stripe *smap, int mirror_num)
487{
488 if (!bioc) {
489 /* Single mirror read/write fast path. */
490 btrfs_bio(bio)->mirror_num = mirror_num;
491 bio->bi_iter.bi_sector = smap->physical >> SECTOR_SHIFT;
492 if (bio_op(bio) != REQ_OP_READ)
493 btrfs_bio(bio)->orig_physical = smap->physical;
494 bio->bi_private = smap->dev;
495 bio->bi_end_io = btrfs_simple_end_io;
496 btrfs_submit_dev_bio(smap->dev, bio);
497 } else if (bioc->map_type & BTRFS_BLOCK_GROUP_RAID56_MASK) {
498 /* Parity RAID write or read recovery. */
499 bio->bi_private = bioc;
500 bio->bi_end_io = btrfs_raid56_end_io;
501 if (bio_op(bio) == REQ_OP_READ)
502 raid56_parity_recover(bio, bioc, mirror_num);
503 else
504 raid56_parity_write(bio, bioc);
505 } else {
506 /* Write to multiple mirrors. */
507 int total_devs = bioc->num_stripes;
508
509 bioc->orig_bio = bio;
510 for (int dev_nr = 0; dev_nr < total_devs; dev_nr++)
511 btrfs_submit_mirrored_bio(bioc, dev_nr);
512 }
513}
514
515static blk_status_t btrfs_bio_csum(struct btrfs_bio *bbio)
516{
517 if (bbio->bio.bi_opf & REQ_META)
518 return btree_csum_one_bio(bbio);
519 return btrfs_csum_one_bio(bbio);
520}
521
522/*
523 * Async submit bios are used to offload expensive checksumming onto the worker
524 * threads.
525 */
526struct async_submit_bio {
527 struct btrfs_bio *bbio;
528 struct btrfs_io_context *bioc;
529 struct btrfs_io_stripe smap;
530 int mirror_num;
531 struct btrfs_work work;
532};
533
534/*
535 * In order to insert checksums into the metadata in large chunks, we wait
536 * until bio submission time. All the pages in the bio are checksummed and
537 * sums are attached onto the ordered extent record.
538 *
539 * At IO completion time the csums attached on the ordered extent record are
540 * inserted into the btree.
541 */
542static void run_one_async_start(struct btrfs_work *work)
543{
544 struct async_submit_bio *async =
545 container_of(work, struct async_submit_bio, work);
546 blk_status_t ret;
547
548 ret = btrfs_bio_csum(async->bbio);
549 if (ret)
550 async->bbio->bio.bi_status = ret;
551}
552
553/*
554 * In order to insert checksums into the metadata in large chunks, we wait
555 * until bio submission time. All the pages in the bio are checksummed and
556 * sums are attached onto the ordered extent record.
557 *
558 * At IO completion time the csums attached on the ordered extent record are
559 * inserted into the tree.
560 *
561 * If called with @do_free == true, then it will free the work struct.
562 */
563static void run_one_async_done(struct btrfs_work *work, bool do_free)
564{
565 struct async_submit_bio *async =
566 container_of(work, struct async_submit_bio, work);
567 struct bio *bio = &async->bbio->bio;
568
569 if (do_free) {
570 kfree(container_of(work, struct async_submit_bio, work));
571 return;
572 }
573
574 /* If an error occurred we just want to clean up the bio and move on. */
575 if (bio->bi_status) {
576 btrfs_bio_end_io(async->bbio, async->bbio->bio.bi_status);
577 return;
578 }
579
580 /*
581 * All of the bios that pass through here are from async helpers.
582 * Use REQ_BTRFS_CGROUP_PUNT to issue them from the owning cgroup's
583 * context. This changes nothing when cgroups aren't in use.
584 */
585 bio->bi_opf |= REQ_BTRFS_CGROUP_PUNT;
586 btrfs_submit_bio(bio, async->bioc, &async->smap, async->mirror_num);
587}
588
589static bool should_async_write(struct btrfs_bio *bbio)
590{
591 bool auto_csum_mode = true;
592
593#ifdef CONFIG_BTRFS_EXPERIMENTAL
594 struct btrfs_fs_devices *fs_devices = bbio->fs_info->fs_devices;
595 enum btrfs_offload_csum_mode csum_mode = READ_ONCE(fs_devices->offload_csum_mode);
596
597 if (csum_mode == BTRFS_OFFLOAD_CSUM_FORCE_OFF)
598 return false;
599
600 auto_csum_mode = (csum_mode == BTRFS_OFFLOAD_CSUM_AUTO);
601#endif
602
603 /* Submit synchronously if the checksum implementation is fast. */
604 if (auto_csum_mode && test_bit(BTRFS_FS_CSUM_IMPL_FAST, &bbio->fs_info->flags))
605 return false;
606
607 /*
608 * Try to defer the submission to a workqueue to parallelize the
609 * checksum calculation unless the I/O is issued synchronously.
610 */
611 if (op_is_sync(bbio->bio.bi_opf))
612 return false;
613
614 /* Zoned devices require I/O to be submitted in order. */
615 if ((bbio->bio.bi_opf & REQ_META) && btrfs_is_zoned(bbio->fs_info))
616 return false;
617
618 return true;
619}
620
621/*
622 * Submit bio to an async queue.
623 *
624 * Return true if the work has been successfully submitted, else false.
625 */
626static bool btrfs_wq_submit_bio(struct btrfs_bio *bbio,
627 struct btrfs_io_context *bioc,
628 struct btrfs_io_stripe *smap, int mirror_num)
629{
630 struct btrfs_fs_info *fs_info = bbio->fs_info;
631 struct async_submit_bio *async;
632
633 async = kmalloc(sizeof(*async), GFP_NOFS);
634 if (!async)
635 return false;
636
637 async->bbio = bbio;
638 async->bioc = bioc;
639 async->smap = *smap;
640 async->mirror_num = mirror_num;
641
642 btrfs_init_work(&async->work, run_one_async_start, run_one_async_done);
643 btrfs_queue_work(fs_info->workers, &async->work);
644 return true;
645}
646
647static u64 btrfs_append_map_length(struct btrfs_bio *bbio, u64 map_length)
648{
649 unsigned int nr_segs;
650 int sector_offset;
651
652 map_length = min(map_length, bbio->fs_info->max_zone_append_size);
653 sector_offset = bio_split_rw_at(&bbio->bio, &bbio->fs_info->limits,
654 &nr_segs, map_length);
655 if (sector_offset) {
656 /*
657 * bio_split_rw_at() could split at a size smaller than our
658 * sectorsize and thus cause unaligned I/Os. Fix that by
659 * always rounding down to the nearest boundary.
660 */
661 return ALIGN_DOWN(sector_offset << SECTOR_SHIFT, bbio->fs_info->sectorsize);
662 }
663 return map_length;
664}
665
666static bool btrfs_submit_chunk(struct btrfs_bio *bbio, int mirror_num)
667{
668 struct btrfs_inode *inode = bbio->inode;
669 struct btrfs_fs_info *fs_info = bbio->fs_info;
670 struct bio *bio = &bbio->bio;
671 u64 logical = bio->bi_iter.bi_sector << SECTOR_SHIFT;
672 u64 length = bio->bi_iter.bi_size;
673 u64 map_length = length;
674 bool use_append = btrfs_use_zone_append(bbio);
675 struct btrfs_io_context *bioc = NULL;
676 struct btrfs_io_stripe smap;
677 blk_status_t ret;
678 int error;
679
680 if (!bbio->inode || btrfs_is_data_reloc_root(inode->root))
681 smap.rst_search_commit_root = true;
682 else
683 smap.rst_search_commit_root = false;
684
685 btrfs_bio_counter_inc_blocked(fs_info);
686 error = btrfs_map_block(fs_info, btrfs_op(bio), logical, &map_length,
687 &bioc, &smap, &mirror_num);
688 if (error) {
689 ret = errno_to_blk_status(error);
690 btrfs_bio_counter_dec(fs_info);
691 goto end_bbio;
692 }
693
694 map_length = min(map_length, length);
695 if (use_append)
696 map_length = btrfs_append_map_length(bbio, map_length);
697
698 if (map_length < length) {
699 struct btrfs_bio *split;
700
701 split = btrfs_split_bio(fs_info, bbio, map_length);
702 if (IS_ERR(split)) {
703 ret = errno_to_blk_status(PTR_ERR(split));
704 btrfs_bio_counter_dec(fs_info);
705 goto end_bbio;
706 }
707 bbio = split;
708 bio = &bbio->bio;
709 }
710
711 /*
712 * Save the iter for the end_io handler and preload the checksums for
713 * data reads.
714 */
715 if (bio_op(bio) == REQ_OP_READ && is_data_bbio(bbio)) {
716 bbio->saved_iter = bio->bi_iter;
717 ret = btrfs_lookup_bio_sums(bbio);
718 if (ret)
719 goto fail;
720 }
721
722 if (btrfs_op(bio) == BTRFS_MAP_WRITE) {
723 if (use_append) {
724 bio->bi_opf &= ~REQ_OP_WRITE;
725 bio->bi_opf |= REQ_OP_ZONE_APPEND;
726 }
727
728 if (is_data_bbio(bbio) && bioc &&
729 btrfs_need_stripe_tree_update(bioc->fs_info, bioc->map_type)) {
730 /*
731 * No locking for the list update, as we only add to
732 * the list in the I/O submission path, and list
733 * iteration only happens in the completion path, which
734 * can't happen until after the last submission.
735 */
736 btrfs_get_bioc(bioc);
737 list_add_tail(&bioc->rst_ordered_entry, &bbio->ordered->bioc_list);
738 }
739
740 /*
741 * Csum items for reloc roots have already been cloned at this
742 * point, so they are handled as part of the no-checksum case.
743 */
744 if (inode && !(inode->flags & BTRFS_INODE_NODATASUM) &&
745 !test_bit(BTRFS_FS_STATE_NO_DATA_CSUMS, &fs_info->fs_state) &&
746 !btrfs_is_data_reloc_root(inode->root)) {
747 if (should_async_write(bbio) &&
748 btrfs_wq_submit_bio(bbio, bioc, &smap, mirror_num))
749 goto done;
750
751 ret = btrfs_bio_csum(bbio);
752 if (ret)
753 goto fail;
754 } else if (use_append ||
755 (btrfs_is_zoned(fs_info) && inode &&
756 inode->flags & BTRFS_INODE_NODATASUM)) {
757 ret = btrfs_alloc_dummy_sum(bbio);
758 if (ret)
759 goto fail;
760 }
761 }
762
763 btrfs_submit_bio(bio, bioc, &smap, mirror_num);
764done:
765 return map_length == length;
766
767fail:
768 btrfs_bio_counter_dec(fs_info);
769 /*
770 * We have split the original bbio, now we have to end both the current
771 * @bbio and remaining one, as the remaining one will never be submitted.
772 */
773 if (map_length < length) {
774 struct btrfs_bio *remaining = bbio->private;
775
776 ASSERT(bbio->bio.bi_pool == &btrfs_clone_bioset);
777 ASSERT(remaining);
778
779 btrfs_bio_end_io(remaining, ret);
780 }
781end_bbio:
782 btrfs_bio_end_io(bbio, ret);
783 /* Do not submit another chunk */
784 return true;
785}
786
787void btrfs_submit_bbio(struct btrfs_bio *bbio, int mirror_num)
788{
789 /* If bbio->inode is not populated, its file_offset must be 0. */
790 ASSERT(bbio->inode || bbio->file_offset == 0);
791
792 while (!btrfs_submit_chunk(bbio, mirror_num))
793 ;
794}
795
796/*
797 * Submit a repair write.
798 *
799 * This bypasses btrfs_submit_bbio() deliberately, as that writes all copies in a
800 * RAID setup. Here we only want to write the one bad copy, so we do the
801 * mapping ourselves and submit the bio directly.
802 *
803 * The I/O is issued synchronously to block the repair read completion from
804 * freeing the bio.
805 */
806int btrfs_repair_io_failure(struct btrfs_fs_info *fs_info, u64 ino, u64 start,
807 u64 length, u64 logical, struct folio *folio,
808 unsigned int folio_offset, int mirror_num)
809{
810 struct btrfs_io_stripe smap = { 0 };
811 struct bio_vec bvec;
812 struct bio bio;
813 int ret = 0;
814
815 ASSERT(!(fs_info->sb->s_flags & SB_RDONLY));
816 BUG_ON(!mirror_num);
817
818 if (btrfs_repair_one_zone(fs_info, logical))
819 return 0;
820
821 /*
822 * Avoid races with device replace and make sure our bioc has devices
823 * associated to its stripes that don't go away while we are doing the
824 * read repair operation.
825 */
826 btrfs_bio_counter_inc_blocked(fs_info);
827 ret = btrfs_map_repair_block(fs_info, &smap, logical, length, mirror_num);
828 if (ret < 0)
829 goto out_counter_dec;
830
831 if (!smap.dev->bdev ||
832 !test_bit(BTRFS_DEV_STATE_WRITEABLE, &smap.dev->dev_state)) {
833 ret = -EIO;
834 goto out_counter_dec;
835 }
836
837 bio_init(&bio, smap.dev->bdev, &bvec, 1, REQ_OP_WRITE | REQ_SYNC);
838 bio.bi_iter.bi_sector = smap.physical >> SECTOR_SHIFT;
839 ret = bio_add_folio(&bio, folio, length, folio_offset);
840 ASSERT(ret);
841 ret = submit_bio_wait(&bio);
842 if (ret) {
843 /* try to remap that extent elsewhere? */
844 btrfs_dev_stat_inc_and_print(smap.dev, BTRFS_DEV_STAT_WRITE_ERRS);
845 goto out_bio_uninit;
846 }
847
848 btrfs_info_rl_in_rcu(fs_info,
849 "read error corrected: ino %llu off %llu (dev %s sector %llu)",
850 ino, start, btrfs_dev_name(smap.dev),
851 smap.physical >> SECTOR_SHIFT);
852 ret = 0;
853
854out_bio_uninit:
855 bio_uninit(&bio);
856out_counter_dec:
857 btrfs_bio_counter_dec(fs_info);
858 return ret;
859}
860
861/*
862 * Submit a btrfs_bio based repair write.
863 *
864 * If @dev_replace is true, the write would be submitted to dev-replace target.
865 */
866void btrfs_submit_repair_write(struct btrfs_bio *bbio, int mirror_num, bool dev_replace)
867{
868 struct btrfs_fs_info *fs_info = bbio->fs_info;
869 u64 logical = bbio->bio.bi_iter.bi_sector << SECTOR_SHIFT;
870 u64 length = bbio->bio.bi_iter.bi_size;
871 struct btrfs_io_stripe smap = { 0 };
872 int ret;
873
874 ASSERT(fs_info);
875 ASSERT(mirror_num > 0);
876 ASSERT(btrfs_op(&bbio->bio) == BTRFS_MAP_WRITE);
877 ASSERT(!bbio->inode);
878
879 btrfs_bio_counter_inc_blocked(fs_info);
880 ret = btrfs_map_repair_block(fs_info, &smap, logical, length, mirror_num);
881 if (ret < 0)
882 goto fail;
883
884 if (dev_replace) {
885 ASSERT(smap.dev == fs_info->dev_replace.srcdev);
886 smap.dev = fs_info->dev_replace.tgtdev;
887 }
888 btrfs_submit_bio(&bbio->bio, NULL, &smap, mirror_num);
889 return;
890
891fail:
892 btrfs_bio_counter_dec(fs_info);
893 btrfs_bio_end_io(bbio, errno_to_blk_status(ret));
894}
895
896int __init btrfs_bioset_init(void)
897{
898 if (bioset_init(&btrfs_bioset, BIO_POOL_SIZE,
899 offsetof(struct btrfs_bio, bio),
900 BIOSET_NEED_BVECS))
901 return -ENOMEM;
902 if (bioset_init(&btrfs_clone_bioset, BIO_POOL_SIZE,
903 offsetof(struct btrfs_bio, bio), 0))
904 goto out_free_bioset;
905 if (bioset_init(&btrfs_repair_bioset, BIO_POOL_SIZE,
906 offsetof(struct btrfs_bio, bio),
907 BIOSET_NEED_BVECS))
908 goto out_free_clone_bioset;
909 if (mempool_init_kmalloc_pool(&btrfs_failed_bio_pool, BIO_POOL_SIZE,
910 sizeof(struct btrfs_failed_bio)))
911 goto out_free_repair_bioset;
912 return 0;
913
914out_free_repair_bioset:
915 bioset_exit(&btrfs_repair_bioset);
916out_free_clone_bioset:
917 bioset_exit(&btrfs_clone_bioset);
918out_free_bioset:
919 bioset_exit(&btrfs_bioset);
920 return -ENOMEM;
921}
922
923void __cold btrfs_bioset_exit(void)
924{
925 mempool_exit(&btrfs_failed_bio_pool);
926 bioset_exit(&btrfs_repair_bioset);
927 bioset_exit(&btrfs_clone_bioset);
928 bioset_exit(&btrfs_bioset);
929}
1// SPDX-License-Identifier: GPL-2.0
2/*
3 * Copyright (C) 2007 Oracle. All rights reserved.
4 * Copyright (C) 2022 Christoph Hellwig.
5 */
6
7#include <linux/bio.h>
8#include "bio.h"
9#include "ctree.h"
10#include "volumes.h"
11#include "raid56.h"
12#include "async-thread.h"
13#include "dev-replace.h"
14#include "zoned.h"
15#include "file-item.h"
16#include "raid-stripe-tree.h"
17
18static struct bio_set btrfs_bioset;
19static struct bio_set btrfs_clone_bioset;
20static struct bio_set btrfs_repair_bioset;
21static mempool_t btrfs_failed_bio_pool;
22
23struct btrfs_failed_bio {
24 struct btrfs_bio *bbio;
25 int num_copies;
26 atomic_t repair_count;
27};
28
29/* Is this a data path I/O that needs storage layer checksum and repair? */
30static inline bool is_data_bbio(struct btrfs_bio *bbio)
31{
32 return bbio->inode && is_data_inode(&bbio->inode->vfs_inode);
33}
34
35static bool bbio_has_ordered_extent(struct btrfs_bio *bbio)
36{
37 return is_data_bbio(bbio) && btrfs_op(&bbio->bio) == BTRFS_MAP_WRITE;
38}
39
40/*
41 * Initialize a btrfs_bio structure. This skips the embedded bio itself as it
42 * is already initialized by the block layer.
43 */
44void btrfs_bio_init(struct btrfs_bio *bbio, struct btrfs_fs_info *fs_info,
45 btrfs_bio_end_io_t end_io, void *private)
46{
47 memset(bbio, 0, offsetof(struct btrfs_bio, bio));
48 bbio->fs_info = fs_info;
49 bbio->end_io = end_io;
50 bbio->private = private;
51 atomic_set(&bbio->pending_ios, 1);
52}
53
54/*
55 * Allocate a btrfs_bio structure. The btrfs_bio is the main I/O container for
56 * btrfs, and is used for all I/O submitted through btrfs_submit_bio.
57 *
58 * Just like the underlying bio_alloc_bioset it will not fail as it is backed by
59 * a mempool.
60 */
61struct btrfs_bio *btrfs_bio_alloc(unsigned int nr_vecs, blk_opf_t opf,
62 struct btrfs_fs_info *fs_info,
63 btrfs_bio_end_io_t end_io, void *private)
64{
65 struct btrfs_bio *bbio;
66 struct bio *bio;
67
68 bio = bio_alloc_bioset(NULL, nr_vecs, opf, GFP_NOFS, &btrfs_bioset);
69 bbio = btrfs_bio(bio);
70 btrfs_bio_init(bbio, fs_info, end_io, private);
71 return bbio;
72}
73
74static struct btrfs_bio *btrfs_split_bio(struct btrfs_fs_info *fs_info,
75 struct btrfs_bio *orig_bbio,
76 u64 map_length, bool use_append)
77{
78 struct btrfs_bio *bbio;
79 struct bio *bio;
80
81 if (use_append) {
82 unsigned int nr_segs;
83
84 bio = bio_split_rw(&orig_bbio->bio, &fs_info->limits, &nr_segs,
85 &btrfs_clone_bioset, map_length);
86 } else {
87 bio = bio_split(&orig_bbio->bio, map_length >> SECTOR_SHIFT,
88 GFP_NOFS, &btrfs_clone_bioset);
89 }
90 bbio = btrfs_bio(bio);
91 btrfs_bio_init(bbio, fs_info, NULL, orig_bbio);
92 bbio->inode = orig_bbio->inode;
93 bbio->file_offset = orig_bbio->file_offset;
94 orig_bbio->file_offset += map_length;
95 if (bbio_has_ordered_extent(bbio)) {
96 refcount_inc(&orig_bbio->ordered->refs);
97 bbio->ordered = orig_bbio->ordered;
98 }
99 atomic_inc(&orig_bbio->pending_ios);
100 return bbio;
101}
102
103/* Free a bio that was never submitted to the underlying device. */
104static void btrfs_cleanup_bio(struct btrfs_bio *bbio)
105{
106 if (bbio_has_ordered_extent(bbio))
107 btrfs_put_ordered_extent(bbio->ordered);
108 bio_put(&bbio->bio);
109}
110
111static void __btrfs_bio_end_io(struct btrfs_bio *bbio)
112{
113 if (bbio_has_ordered_extent(bbio)) {
114 struct btrfs_ordered_extent *ordered = bbio->ordered;
115
116 bbio->end_io(bbio);
117 btrfs_put_ordered_extent(ordered);
118 } else {
119 bbio->end_io(bbio);
120 }
121}
122
123void btrfs_bio_end_io(struct btrfs_bio *bbio, blk_status_t status)
124{
125 bbio->bio.bi_status = status;
126 __btrfs_bio_end_io(bbio);
127}
128
129static void btrfs_orig_write_end_io(struct bio *bio);
130
131static void btrfs_bbio_propagate_error(struct btrfs_bio *bbio,
132 struct btrfs_bio *orig_bbio)
133{
134 /*
135 * For writes we tolerate nr_mirrors - 1 write failures, so we can't
136 * just blindly propagate a write failure here. Instead increment the
137 * error count in the original I/O context so that it is guaranteed to
138 * be larger than the error tolerance.
139 */
140 if (bbio->bio.bi_end_io == &btrfs_orig_write_end_io) {
141 struct btrfs_io_stripe *orig_stripe = orig_bbio->bio.bi_private;
142 struct btrfs_io_context *orig_bioc = orig_stripe->bioc;
143
144 atomic_add(orig_bioc->max_errors, &orig_bioc->error);
145 } else {
146 orig_bbio->bio.bi_status = bbio->bio.bi_status;
147 }
148}
149
150static void btrfs_orig_bbio_end_io(struct btrfs_bio *bbio)
151{
152 if (bbio->bio.bi_pool == &btrfs_clone_bioset) {
153 struct btrfs_bio *orig_bbio = bbio->private;
154
155 if (bbio->bio.bi_status)
156 btrfs_bbio_propagate_error(bbio, orig_bbio);
157 btrfs_cleanup_bio(bbio);
158 bbio = orig_bbio;
159 }
160
161 if (atomic_dec_and_test(&bbio->pending_ios))
162 __btrfs_bio_end_io(bbio);
163}
164
165static int next_repair_mirror(struct btrfs_failed_bio *fbio, int cur_mirror)
166{
167 if (cur_mirror == fbio->num_copies)
168 return cur_mirror + 1 - fbio->num_copies;
169 return cur_mirror + 1;
170}
171
172static int prev_repair_mirror(struct btrfs_failed_bio *fbio, int cur_mirror)
173{
174 if (cur_mirror == 1)
175 return fbio->num_copies;
176 return cur_mirror - 1;
177}
178
179static void btrfs_repair_done(struct btrfs_failed_bio *fbio)
180{
181 if (atomic_dec_and_test(&fbio->repair_count)) {
182 btrfs_orig_bbio_end_io(fbio->bbio);
183 mempool_free(fbio, &btrfs_failed_bio_pool);
184 }
185}
186
187static void btrfs_end_repair_bio(struct btrfs_bio *repair_bbio,
188 struct btrfs_device *dev)
189{
190 struct btrfs_failed_bio *fbio = repair_bbio->private;
191 struct btrfs_inode *inode = repair_bbio->inode;
192 struct btrfs_fs_info *fs_info = inode->root->fs_info;
193 struct bio_vec *bv = bio_first_bvec_all(&repair_bbio->bio);
194 int mirror = repair_bbio->mirror_num;
195
196 /*
197 * We can only trigger this for data bio, which doesn't support larger
198 * folios yet.
199 */
200 ASSERT(folio_order(page_folio(bv->bv_page)) == 0);
201
202 if (repair_bbio->bio.bi_status ||
203 !btrfs_data_csum_ok(repair_bbio, dev, 0, bv)) {
204 bio_reset(&repair_bbio->bio, NULL, REQ_OP_READ);
205 repair_bbio->bio.bi_iter = repair_bbio->saved_iter;
206
207 mirror = next_repair_mirror(fbio, mirror);
208 if (mirror == fbio->bbio->mirror_num) {
209 btrfs_debug(fs_info, "no mirror left");
210 fbio->bbio->bio.bi_status = BLK_STS_IOERR;
211 goto done;
212 }
213
214 btrfs_submit_bio(repair_bbio, mirror);
215 return;
216 }
217
218 do {
219 mirror = prev_repair_mirror(fbio, mirror);
220 btrfs_repair_io_failure(fs_info, btrfs_ino(inode),
221 repair_bbio->file_offset, fs_info->sectorsize,
222 repair_bbio->saved_iter.bi_sector << SECTOR_SHIFT,
223 page_folio(bv->bv_page), bv->bv_offset, mirror);
224 } while (mirror != fbio->bbio->mirror_num);
225
226done:
227 btrfs_repair_done(fbio);
228 bio_put(&repair_bbio->bio);
229}
230
231/*
232 * Try to kick off a repair read to the next available mirror for a bad sector.
233 *
234 * This primarily tries to recover good data to serve the actual read request,
235 * but also tries to write the good data back to the bad mirror(s) when a
236 * read succeeded to restore the redundancy.
237 */
238static struct btrfs_failed_bio *repair_one_sector(struct btrfs_bio *failed_bbio,
239 u32 bio_offset,
240 struct bio_vec *bv,
241 struct btrfs_failed_bio *fbio)
242{
243 struct btrfs_inode *inode = failed_bbio->inode;
244 struct btrfs_fs_info *fs_info = inode->root->fs_info;
245 const u32 sectorsize = fs_info->sectorsize;
246 const u64 logical = (failed_bbio->saved_iter.bi_sector << SECTOR_SHIFT);
247 struct btrfs_bio *repair_bbio;
248 struct bio *repair_bio;
249 int num_copies;
250 int mirror;
251
252 btrfs_debug(fs_info, "repair read error: read error at %llu",
253 failed_bbio->file_offset + bio_offset);
254
255 num_copies = btrfs_num_copies(fs_info, logical, sectorsize);
256 if (num_copies == 1) {
257 btrfs_debug(fs_info, "no copy to repair from");
258 failed_bbio->bio.bi_status = BLK_STS_IOERR;
259 return fbio;
260 }
261
262 if (!fbio) {
263 fbio = mempool_alloc(&btrfs_failed_bio_pool, GFP_NOFS);
264 fbio->bbio = failed_bbio;
265 fbio->num_copies = num_copies;
266 atomic_set(&fbio->repair_count, 1);
267 }
268
269 atomic_inc(&fbio->repair_count);
270
271 repair_bio = bio_alloc_bioset(NULL, 1, REQ_OP_READ, GFP_NOFS,
272 &btrfs_repair_bioset);
273 repair_bio->bi_iter.bi_sector = failed_bbio->saved_iter.bi_sector;
274 __bio_add_page(repair_bio, bv->bv_page, bv->bv_len, bv->bv_offset);
275
276 repair_bbio = btrfs_bio(repair_bio);
277 btrfs_bio_init(repair_bbio, fs_info, NULL, fbio);
278 repair_bbio->inode = failed_bbio->inode;
279 repair_bbio->file_offset = failed_bbio->file_offset + bio_offset;
280
281 mirror = next_repair_mirror(fbio, failed_bbio->mirror_num);
282 btrfs_debug(fs_info, "submitting repair read to mirror %d", mirror);
283 btrfs_submit_bio(repair_bbio, mirror);
284 return fbio;
285}
286
287static void btrfs_check_read_bio(struct btrfs_bio *bbio, struct btrfs_device *dev)
288{
289 struct btrfs_inode *inode = bbio->inode;
290 struct btrfs_fs_info *fs_info = inode->root->fs_info;
291 u32 sectorsize = fs_info->sectorsize;
292 struct bvec_iter *iter = &bbio->saved_iter;
293 blk_status_t status = bbio->bio.bi_status;
294 struct btrfs_failed_bio *fbio = NULL;
295 u32 offset = 0;
296
297 /* Read-repair requires the inode field to be set by the submitter. */
298 ASSERT(inode);
299
300 /*
301 * Hand off repair bios to the repair code as there is no upper level
302 * submitter for them.
303 */
304 if (bbio->bio.bi_pool == &btrfs_repair_bioset) {
305 btrfs_end_repair_bio(bbio, dev);
306 return;
307 }
308
309 /* Clear the I/O error. A failed repair will reset it. */
310 bbio->bio.bi_status = BLK_STS_OK;
311
312 while (iter->bi_size) {
313 struct bio_vec bv = bio_iter_iovec(&bbio->bio, *iter);
314
315 bv.bv_len = min(bv.bv_len, sectorsize);
316 if (status || !btrfs_data_csum_ok(bbio, dev, offset, &bv))
317 fbio = repair_one_sector(bbio, offset, &bv, fbio);
318
319 bio_advance_iter_single(&bbio->bio, iter, sectorsize);
320 offset += sectorsize;
321 }
322
323 if (bbio->csum != bbio->csum_inline)
324 kfree(bbio->csum);
325
326 if (fbio)
327 btrfs_repair_done(fbio);
328 else
329 btrfs_orig_bbio_end_io(bbio);
330}
331
332static void btrfs_log_dev_io_error(struct bio *bio, struct btrfs_device *dev)
333{
334 if (!dev || !dev->bdev)
335 return;
336 if (bio->bi_status != BLK_STS_IOERR && bio->bi_status != BLK_STS_TARGET)
337 return;
338
339 if (btrfs_op(bio) == BTRFS_MAP_WRITE)
340 btrfs_dev_stat_inc_and_print(dev, BTRFS_DEV_STAT_WRITE_ERRS);
341 else if (!(bio->bi_opf & REQ_RAHEAD))
342 btrfs_dev_stat_inc_and_print(dev, BTRFS_DEV_STAT_READ_ERRS);
343 if (bio->bi_opf & REQ_PREFLUSH)
344 btrfs_dev_stat_inc_and_print(dev, BTRFS_DEV_STAT_FLUSH_ERRS);
345}
346
347static struct workqueue_struct *btrfs_end_io_wq(struct btrfs_fs_info *fs_info,
348 struct bio *bio)
349{
350 if (bio->bi_opf & REQ_META)
351 return fs_info->endio_meta_workers;
352 return fs_info->endio_workers;
353}
354
355static void btrfs_end_bio_work(struct work_struct *work)
356{
357 struct btrfs_bio *bbio = container_of(work, struct btrfs_bio, end_io_work);
358
359 /* Metadata reads are checked and repaired by the submitter. */
360 if (is_data_bbio(bbio))
361 btrfs_check_read_bio(bbio, bbio->bio.bi_private);
362 else
363 btrfs_orig_bbio_end_io(bbio);
364}
365
366static void btrfs_simple_end_io(struct bio *bio)
367{
368 struct btrfs_bio *bbio = btrfs_bio(bio);
369 struct btrfs_device *dev = bio->bi_private;
370 struct btrfs_fs_info *fs_info = bbio->fs_info;
371
372 btrfs_bio_counter_dec(fs_info);
373
374 if (bio->bi_status)
375 btrfs_log_dev_io_error(bio, dev);
376
377 if (bio_op(bio) == REQ_OP_READ) {
378 INIT_WORK(&bbio->end_io_work, btrfs_end_bio_work);
379 queue_work(btrfs_end_io_wq(fs_info, bio), &bbio->end_io_work);
380 } else {
381 if (bio_op(bio) == REQ_OP_ZONE_APPEND && !bio->bi_status)
382 btrfs_record_physical_zoned(bbio);
383 btrfs_orig_bbio_end_io(bbio);
384 }
385}
386
387static void btrfs_raid56_end_io(struct bio *bio)
388{
389 struct btrfs_io_context *bioc = bio->bi_private;
390 struct btrfs_bio *bbio = btrfs_bio(bio);
391
392 btrfs_bio_counter_dec(bioc->fs_info);
393 bbio->mirror_num = bioc->mirror_num;
394 if (bio_op(bio) == REQ_OP_READ && is_data_bbio(bbio))
395 btrfs_check_read_bio(bbio, NULL);
396 else
397 btrfs_orig_bbio_end_io(bbio);
398
399 btrfs_put_bioc(bioc);
400}
401
402static void btrfs_orig_write_end_io(struct bio *bio)
403{
404 struct btrfs_io_stripe *stripe = bio->bi_private;
405 struct btrfs_io_context *bioc = stripe->bioc;
406 struct btrfs_bio *bbio = btrfs_bio(bio);
407
408 btrfs_bio_counter_dec(bioc->fs_info);
409
410 if (bio->bi_status) {
411 atomic_inc(&bioc->error);
412 btrfs_log_dev_io_error(bio, stripe->dev);
413 }
414
415 /*
416 * Only send an error to the higher layers if it is beyond the tolerance
417 * threshold.
418 */
419 if (atomic_read(&bioc->error) > bioc->max_errors)
420 bio->bi_status = BLK_STS_IOERR;
421 else
422 bio->bi_status = BLK_STS_OK;
423
424 if (bio_op(bio) == REQ_OP_ZONE_APPEND && !bio->bi_status)
425 stripe->physical = bio->bi_iter.bi_sector << SECTOR_SHIFT;
426
427 btrfs_orig_bbio_end_io(bbio);
428 btrfs_put_bioc(bioc);
429}
430
431static void btrfs_clone_write_end_io(struct bio *bio)
432{
433 struct btrfs_io_stripe *stripe = bio->bi_private;
434
435 if (bio->bi_status) {
436 atomic_inc(&stripe->bioc->error);
437 btrfs_log_dev_io_error(bio, stripe->dev);
438 } else if (bio_op(bio) == REQ_OP_ZONE_APPEND) {
439 stripe->physical = bio->bi_iter.bi_sector << SECTOR_SHIFT;
440 }
441
442 /* Pass on control to the original bio this one was cloned from */
443 bio_endio(stripe->bioc->orig_bio);
444 bio_put(bio);
445}
446
447static void btrfs_submit_dev_bio(struct btrfs_device *dev, struct bio *bio)
448{
449 if (!dev || !dev->bdev ||
450 test_bit(BTRFS_DEV_STATE_MISSING, &dev->dev_state) ||
451 (btrfs_op(bio) == BTRFS_MAP_WRITE &&
452 !test_bit(BTRFS_DEV_STATE_WRITEABLE, &dev->dev_state))) {
453 bio_io_error(bio);
454 return;
455 }
456
457 bio_set_dev(bio, dev->bdev);
458
459 /*
460 * For zone append writing, bi_sector must point the beginning of the
461 * zone
462 */
463 if (bio_op(bio) == REQ_OP_ZONE_APPEND) {
464 u64 physical = bio->bi_iter.bi_sector << SECTOR_SHIFT;
465 u64 zone_start = round_down(physical, dev->fs_info->zone_size);
466
467 ASSERT(btrfs_dev_is_sequential(dev, physical));
468 bio->bi_iter.bi_sector = zone_start >> SECTOR_SHIFT;
469 }
470 btrfs_debug_in_rcu(dev->fs_info,
471 "%s: rw %d 0x%x, sector=%llu, dev=%lu (%s id %llu), size=%u",
472 __func__, bio_op(bio), bio->bi_opf, bio->bi_iter.bi_sector,
473 (unsigned long)dev->bdev->bd_dev, btrfs_dev_name(dev),
474 dev->devid, bio->bi_iter.bi_size);
475
476 if (bio->bi_opf & REQ_BTRFS_CGROUP_PUNT)
477 blkcg_punt_bio_submit(bio);
478 else
479 submit_bio(bio);
480}
481
482static void btrfs_submit_mirrored_bio(struct btrfs_io_context *bioc, int dev_nr)
483{
484 struct bio *orig_bio = bioc->orig_bio, *bio;
485
486 ASSERT(bio_op(orig_bio) != REQ_OP_READ);
487
488 /* Reuse the bio embedded into the btrfs_bio for the last mirror */
489 if (dev_nr == bioc->num_stripes - 1) {
490 bio = orig_bio;
491 bio->bi_end_io = btrfs_orig_write_end_io;
492 } else {
493 bio = bio_alloc_clone(NULL, orig_bio, GFP_NOFS, &fs_bio_set);
494 bio_inc_remaining(orig_bio);
495 bio->bi_end_io = btrfs_clone_write_end_io;
496 }
497
498 bio->bi_private = &bioc->stripes[dev_nr];
499 bio->bi_iter.bi_sector = bioc->stripes[dev_nr].physical >> SECTOR_SHIFT;
500 bioc->stripes[dev_nr].bioc = bioc;
501 bioc->size = bio->bi_iter.bi_size;
502 btrfs_submit_dev_bio(bioc->stripes[dev_nr].dev, bio);
503}
504
505static void __btrfs_submit_bio(struct bio *bio, struct btrfs_io_context *bioc,
506 struct btrfs_io_stripe *smap, int mirror_num)
507{
508 if (!bioc) {
509 /* Single mirror read/write fast path. */
510 btrfs_bio(bio)->mirror_num = mirror_num;
511 bio->bi_iter.bi_sector = smap->physical >> SECTOR_SHIFT;
512 if (bio_op(bio) != REQ_OP_READ)
513 btrfs_bio(bio)->orig_physical = smap->physical;
514 bio->bi_private = smap->dev;
515 bio->bi_end_io = btrfs_simple_end_io;
516 btrfs_submit_dev_bio(smap->dev, bio);
517 } else if (bioc->map_type & BTRFS_BLOCK_GROUP_RAID56_MASK) {
518 /* Parity RAID write or read recovery. */
519 bio->bi_private = bioc;
520 bio->bi_end_io = btrfs_raid56_end_io;
521 if (bio_op(bio) == REQ_OP_READ)
522 raid56_parity_recover(bio, bioc, mirror_num);
523 else
524 raid56_parity_write(bio, bioc);
525 } else {
526 /* Write to multiple mirrors. */
527 int total_devs = bioc->num_stripes;
528
529 bioc->orig_bio = bio;
530 for (int dev_nr = 0; dev_nr < total_devs; dev_nr++)
531 btrfs_submit_mirrored_bio(bioc, dev_nr);
532 }
533}
534
535static blk_status_t btrfs_bio_csum(struct btrfs_bio *bbio)
536{
537 if (bbio->bio.bi_opf & REQ_META)
538 return btree_csum_one_bio(bbio);
539 return btrfs_csum_one_bio(bbio);
540}
541
542/*
543 * Async submit bios are used to offload expensive checksumming onto the worker
544 * threads.
545 */
546struct async_submit_bio {
547 struct btrfs_bio *bbio;
548 struct btrfs_io_context *bioc;
549 struct btrfs_io_stripe smap;
550 int mirror_num;
551 struct btrfs_work work;
552};
553
554/*
555 * In order to insert checksums into the metadata in large chunks, we wait
556 * until bio submission time. All the pages in the bio are checksummed and
557 * sums are attached onto the ordered extent record.
558 *
559 * At IO completion time the csums attached on the ordered extent record are
560 * inserted into the btree.
561 */
562static void run_one_async_start(struct btrfs_work *work)
563{
564 struct async_submit_bio *async =
565 container_of(work, struct async_submit_bio, work);
566 blk_status_t ret;
567
568 ret = btrfs_bio_csum(async->bbio);
569 if (ret)
570 async->bbio->bio.bi_status = ret;
571}
572
573/*
574 * In order to insert checksums into the metadata in large chunks, we wait
575 * until bio submission time. All the pages in the bio are checksummed and
576 * sums are attached onto the ordered extent record.
577 *
578 * At IO completion time the csums attached on the ordered extent record are
579 * inserted into the tree.
580 *
581 * If called with @do_free == true, then it will free the work struct.
582 */
583static void run_one_async_done(struct btrfs_work *work, bool do_free)
584{
585 struct async_submit_bio *async =
586 container_of(work, struct async_submit_bio, work);
587 struct bio *bio = &async->bbio->bio;
588
589 if (do_free) {
590 kfree(container_of(work, struct async_submit_bio, work));
591 return;
592 }
593
594 /* If an error occurred we just want to clean up the bio and move on. */
595 if (bio->bi_status) {
596 btrfs_orig_bbio_end_io(async->bbio);
597 return;
598 }
599
600 /*
601 * All of the bios that pass through here are from async helpers.
602 * Use REQ_BTRFS_CGROUP_PUNT to issue them from the owning cgroup's
603 * context. This changes nothing when cgroups aren't in use.
604 */
605 bio->bi_opf |= REQ_BTRFS_CGROUP_PUNT;
606 __btrfs_submit_bio(bio, async->bioc, &async->smap, async->mirror_num);
607}
608
609static bool should_async_write(struct btrfs_bio *bbio)
610{
611 bool auto_csum_mode = true;
612
613#ifdef CONFIG_BTRFS_DEBUG
614 struct btrfs_fs_devices *fs_devices = bbio->fs_info->fs_devices;
615 enum btrfs_offload_csum_mode csum_mode = READ_ONCE(fs_devices->offload_csum_mode);
616
617 if (csum_mode == BTRFS_OFFLOAD_CSUM_FORCE_OFF)
618 return false;
619
620 auto_csum_mode = (csum_mode == BTRFS_OFFLOAD_CSUM_AUTO);
621#endif
622
623 /* Submit synchronously if the checksum implementation is fast. */
624 if (auto_csum_mode && test_bit(BTRFS_FS_CSUM_IMPL_FAST, &bbio->fs_info->flags))
625 return false;
626
627 /*
628 * Try to defer the submission to a workqueue to parallelize the
629 * checksum calculation unless the I/O is issued synchronously.
630 */
631 if (op_is_sync(bbio->bio.bi_opf))
632 return false;
633
634 /* Zoned devices require I/O to be submitted in order. */
635 if ((bbio->bio.bi_opf & REQ_META) && btrfs_is_zoned(bbio->fs_info))
636 return false;
637
638 return true;
639}
640
641/*
642 * Submit bio to an async queue.
643 *
644 * Return true if the work has been successfully submitted, else false.
645 */
646static bool btrfs_wq_submit_bio(struct btrfs_bio *bbio,
647 struct btrfs_io_context *bioc,
648 struct btrfs_io_stripe *smap, int mirror_num)
649{
650 struct btrfs_fs_info *fs_info = bbio->fs_info;
651 struct async_submit_bio *async;
652
653 async = kmalloc(sizeof(*async), GFP_NOFS);
654 if (!async)
655 return false;
656
657 async->bbio = bbio;
658 async->bioc = bioc;
659 async->smap = *smap;
660 async->mirror_num = mirror_num;
661
662 btrfs_init_work(&async->work, run_one_async_start, run_one_async_done);
663 btrfs_queue_work(fs_info->workers, &async->work);
664 return true;
665}
666
667static bool btrfs_submit_chunk(struct btrfs_bio *bbio, int mirror_num)
668{
669 struct btrfs_inode *inode = bbio->inode;
670 struct btrfs_fs_info *fs_info = bbio->fs_info;
671 struct btrfs_bio *orig_bbio = bbio;
672 struct bio *bio = &bbio->bio;
673 u64 logical = bio->bi_iter.bi_sector << SECTOR_SHIFT;
674 u64 length = bio->bi_iter.bi_size;
675 u64 map_length = length;
676 bool use_append = btrfs_use_zone_append(bbio);
677 struct btrfs_io_context *bioc = NULL;
678 struct btrfs_io_stripe smap;
679 blk_status_t ret;
680 int error;
681
682 smap.is_scrub = !bbio->inode;
683
684 btrfs_bio_counter_inc_blocked(fs_info);
685 error = btrfs_map_block(fs_info, btrfs_op(bio), logical, &map_length,
686 &bioc, &smap, &mirror_num);
687 if (error) {
688 ret = errno_to_blk_status(error);
689 goto fail;
690 }
691
692 map_length = min(map_length, length);
693 if (use_append)
694 map_length = min(map_length, fs_info->max_zone_append_size);
695
696 if (map_length < length) {
697 bbio = btrfs_split_bio(fs_info, bbio, map_length, use_append);
698 bio = &bbio->bio;
699 }
700
701 /*
702 * Save the iter for the end_io handler and preload the checksums for
703 * data reads.
704 */
705 if (bio_op(bio) == REQ_OP_READ && is_data_bbio(bbio)) {
706 bbio->saved_iter = bio->bi_iter;
707 ret = btrfs_lookup_bio_sums(bbio);
708 if (ret)
709 goto fail_put_bio;
710 }
711
712 if (btrfs_op(bio) == BTRFS_MAP_WRITE) {
713 if (use_append) {
714 bio->bi_opf &= ~REQ_OP_WRITE;
715 bio->bi_opf |= REQ_OP_ZONE_APPEND;
716 }
717
718 if (is_data_bbio(bbio) && bioc &&
719 btrfs_need_stripe_tree_update(bioc->fs_info, bioc->map_type)) {
720 /*
721 * No locking for the list update, as we only add to
722 * the list in the I/O submission path, and list
723 * iteration only happens in the completion path, which
724 * can't happen until after the last submission.
725 */
726 btrfs_get_bioc(bioc);
727 list_add_tail(&bioc->rst_ordered_entry, &bbio->ordered->bioc_list);
728 }
729
730 /*
731 * Csum items for reloc roots have already been cloned at this
732 * point, so they are handled as part of the no-checksum case.
733 */
734 if (inode && !(inode->flags & BTRFS_INODE_NODATASUM) &&
735 !test_bit(BTRFS_FS_STATE_NO_CSUMS, &fs_info->fs_state) &&
736 !btrfs_is_data_reloc_root(inode->root)) {
737 if (should_async_write(bbio) &&
738 btrfs_wq_submit_bio(bbio, bioc, &smap, mirror_num))
739 goto done;
740
741 ret = btrfs_bio_csum(bbio);
742 if (ret)
743 goto fail_put_bio;
744 } else if (use_append) {
745 ret = btrfs_alloc_dummy_sum(bbio);
746 if (ret)
747 goto fail_put_bio;
748 }
749 }
750
751 __btrfs_submit_bio(bio, bioc, &smap, mirror_num);
752done:
753 return map_length == length;
754
755fail_put_bio:
756 if (map_length < length)
757 btrfs_cleanup_bio(bbio);
758fail:
759 btrfs_bio_counter_dec(fs_info);
760 btrfs_bio_end_io(orig_bbio, ret);
761 /* Do not submit another chunk */
762 return true;
763}
764
765void btrfs_submit_bio(struct btrfs_bio *bbio, int mirror_num)
766{
767 /* If bbio->inode is not populated, its file_offset must be 0. */
768 ASSERT(bbio->inode || bbio->file_offset == 0);
769
770 while (!btrfs_submit_chunk(bbio, mirror_num))
771 ;
772}
773
774/*
775 * Submit a repair write.
776 *
777 * This bypasses btrfs_submit_bio deliberately, as that writes all copies in a
778 * RAID setup. Here we only want to write the one bad copy, so we do the
779 * mapping ourselves and submit the bio directly.
780 *
781 * The I/O is issued synchronously to block the repair read completion from
782 * freeing the bio.
783 */
784int btrfs_repair_io_failure(struct btrfs_fs_info *fs_info, u64 ino, u64 start,
785 u64 length, u64 logical, struct folio *folio,
786 unsigned int folio_offset, int mirror_num)
787{
788 struct btrfs_io_stripe smap = { 0 };
789 struct bio_vec bvec;
790 struct bio bio;
791 int ret = 0;
792
793 ASSERT(!(fs_info->sb->s_flags & SB_RDONLY));
794 BUG_ON(!mirror_num);
795
796 if (btrfs_repair_one_zone(fs_info, logical))
797 return 0;
798
799 /*
800 * Avoid races with device replace and make sure our bioc has devices
801 * associated to its stripes that don't go away while we are doing the
802 * read repair operation.
803 */
804 btrfs_bio_counter_inc_blocked(fs_info);
805 ret = btrfs_map_repair_block(fs_info, &smap, logical, length, mirror_num);
806 if (ret < 0)
807 goto out_counter_dec;
808
809 if (!smap.dev->bdev ||
810 !test_bit(BTRFS_DEV_STATE_WRITEABLE, &smap.dev->dev_state)) {
811 ret = -EIO;
812 goto out_counter_dec;
813 }
814
815 bio_init(&bio, smap.dev->bdev, &bvec, 1, REQ_OP_WRITE | REQ_SYNC);
816 bio.bi_iter.bi_sector = smap.physical >> SECTOR_SHIFT;
817 ret = bio_add_folio(&bio, folio, length, folio_offset);
818 ASSERT(ret);
819 ret = submit_bio_wait(&bio);
820 if (ret) {
821 /* try to remap that extent elsewhere? */
822 btrfs_dev_stat_inc_and_print(smap.dev, BTRFS_DEV_STAT_WRITE_ERRS);
823 goto out_bio_uninit;
824 }
825
826 btrfs_info_rl_in_rcu(fs_info,
827 "read error corrected: ino %llu off %llu (dev %s sector %llu)",
828 ino, start, btrfs_dev_name(smap.dev),
829 smap.physical >> SECTOR_SHIFT);
830 ret = 0;
831
832out_bio_uninit:
833 bio_uninit(&bio);
834out_counter_dec:
835 btrfs_bio_counter_dec(fs_info);
836 return ret;
837}
838
839/*
840 * Submit a btrfs_bio based repair write.
841 *
842 * If @dev_replace is true, the write would be submitted to dev-replace target.
843 */
844void btrfs_submit_repair_write(struct btrfs_bio *bbio, int mirror_num, bool dev_replace)
845{
846 struct btrfs_fs_info *fs_info = bbio->fs_info;
847 u64 logical = bbio->bio.bi_iter.bi_sector << SECTOR_SHIFT;
848 u64 length = bbio->bio.bi_iter.bi_size;
849 struct btrfs_io_stripe smap = { 0 };
850 int ret;
851
852 ASSERT(fs_info);
853 ASSERT(mirror_num > 0);
854 ASSERT(btrfs_op(&bbio->bio) == BTRFS_MAP_WRITE);
855 ASSERT(!bbio->inode);
856
857 btrfs_bio_counter_inc_blocked(fs_info);
858 ret = btrfs_map_repair_block(fs_info, &smap, logical, length, mirror_num);
859 if (ret < 0)
860 goto fail;
861
862 if (dev_replace) {
863 ASSERT(smap.dev == fs_info->dev_replace.srcdev);
864 smap.dev = fs_info->dev_replace.tgtdev;
865 }
866 __btrfs_submit_bio(&bbio->bio, NULL, &smap, mirror_num);
867 return;
868
869fail:
870 btrfs_bio_counter_dec(fs_info);
871 btrfs_bio_end_io(bbio, errno_to_blk_status(ret));
872}
873
874int __init btrfs_bioset_init(void)
875{
876 if (bioset_init(&btrfs_bioset, BIO_POOL_SIZE,
877 offsetof(struct btrfs_bio, bio),
878 BIOSET_NEED_BVECS))
879 return -ENOMEM;
880 if (bioset_init(&btrfs_clone_bioset, BIO_POOL_SIZE,
881 offsetof(struct btrfs_bio, bio), 0))
882 goto out_free_bioset;
883 if (bioset_init(&btrfs_repair_bioset, BIO_POOL_SIZE,
884 offsetof(struct btrfs_bio, bio),
885 BIOSET_NEED_BVECS))
886 goto out_free_clone_bioset;
887 if (mempool_init_kmalloc_pool(&btrfs_failed_bio_pool, BIO_POOL_SIZE,
888 sizeof(struct btrfs_failed_bio)))
889 goto out_free_repair_bioset;
890 return 0;
891
892out_free_repair_bioset:
893 bioset_exit(&btrfs_repair_bioset);
894out_free_clone_bioset:
895 bioset_exit(&btrfs_clone_bioset);
896out_free_bioset:
897 bioset_exit(&btrfs_bioset);
898 return -ENOMEM;
899}
900
901void __cold btrfs_bioset_exit(void)
902{
903 mempool_exit(&btrfs_failed_bio_pool);
904 bioset_exit(&btrfs_repair_bioset);
905 bioset_exit(&btrfs_clone_bioset);
906 bioset_exit(&btrfs_bioset);
907}