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