bio.c

   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(const struct btrfs_bio *bbio)
  31{
  32	return bbio->inode && is_data_inode(bbio->inode);
  33}
  34
  35static bool bbio_has_ordered_extent(const 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_inode *inode, u64 file_offset,
  45		    btrfs_bio_end_io_t end_io, void *private)
  46{
  47	/* @inode parameter is mandatory. */
  48	ASSERT(inode);
  49
  50	memset(bbio, 0, offsetof(struct btrfs_bio, bio));
  51	bbio->inode = inode;
  52	bbio->end_io = end_io;
  53	bbio->private = private;
  54	bbio->file_offset = file_offset;
  55	atomic_set(&bbio->pending_ios, 1);
  56	WRITE_ONCE(bbio->status, BLK_STS_OK);
  57}
  58
  59/*
  60 * Allocate a btrfs_bio structure.  The btrfs_bio is the main I/O container for
  61 * btrfs, and is used for all I/O submitted through btrfs_submit_bbio().
  62 *
  63 * Just like the underlying bio_alloc_bioset it will not fail as it is backed by
  64 * a mempool.
  65 */
  66struct btrfs_bio *btrfs_bio_alloc(unsigned int nr_vecs, blk_opf_t opf,
  67				  struct btrfs_inode *inode, u64 file_offset,
  68				  btrfs_bio_end_io_t end_io, void *private)
  69{
  70	struct btrfs_bio *bbio;
  71	struct bio *bio;
  72
  73	bio = bio_alloc_bioset(NULL, nr_vecs, opf, GFP_NOFS, &btrfs_bioset);
  74	bbio = btrfs_bio(bio);
  75	btrfs_bio_init(bbio, inode, file_offset, end_io, private);
  76	return bbio;
  77}
  78
  79static struct btrfs_bio *btrfs_split_bio(struct btrfs_fs_info *fs_info,
  80					 struct btrfs_bio *orig_bbio,
  81					 u64 map_length)
  82{
  83	struct btrfs_bio *bbio;
  84	struct bio *bio;
  85
  86	bio = bio_split(&orig_bbio->bio, map_length >> SECTOR_SHIFT, GFP_NOFS,
  87			&btrfs_clone_bioset);
  88	if (IS_ERR(bio))
  89		return ERR_CAST(bio);
  90
  91	bbio = btrfs_bio(bio);
  92	btrfs_bio_init(bbio, orig_bbio->inode, orig_bbio->file_offset, NULL, orig_bbio);
  93	orig_bbio->file_offset += map_length;
  94	if (bbio_has_ordered_extent(bbio)) {
  95		refcount_inc(&orig_bbio->ordered->refs);
  96		bbio->ordered = orig_bbio->ordered;
  97		bbio->orig_logical = orig_bbio->orig_logical;
  98		orig_bbio->orig_logical += map_length;
  99	}
 100	bbio->csum_search_commit_root = orig_bbio->csum_search_commit_root;
 101	atomic_inc(&orig_bbio->pending_ios);
 102	return bbio;
 103}
 104
 105void btrfs_bio_end_io(struct btrfs_bio *bbio, blk_status_t status)
 106{
 107	/* Make sure we're already in task context. */
 108	ASSERT(in_task());
 109
 110	if (bbio->async_csum)
 111		wait_for_completion(&bbio->csum_done);
 112
 113	bbio->bio.bi_status = status;
 114	if (bbio->bio.bi_pool == &btrfs_clone_bioset) {
 115		struct btrfs_bio *orig_bbio = bbio->private;
 116
 117		/* Free bio that was never submitted to the underlying device. */
 118		if (bbio_has_ordered_extent(bbio))
 119			btrfs_put_ordered_extent(bbio->ordered);
 120		bio_put(&bbio->bio);
 121
 122		bbio = orig_bbio;
 123	}
 124
 125	/*
 126	 * At this point, bbio always points to the original btrfs_bio. Save
 127	 * the first error in it.
 128	 */
 129	if (status != BLK_STS_OK)
 130		cmpxchg(&bbio->status, BLK_STS_OK, status);
 131
 132	if (atomic_dec_and_test(&bbio->pending_ios)) {
 133		/* Load split bio's error which might be set above. */
 134		if (status == BLK_STS_OK)
 135			bbio->bio.bi_status = READ_ONCE(bbio->status);
 136
 137		if (bbio_has_ordered_extent(bbio)) {
 138			struct btrfs_ordered_extent *ordered = bbio->ordered;
 139
 140			bbio->end_io(bbio);
 141			btrfs_put_ordered_extent(ordered);
 142		} else {
 143			bbio->end_io(bbio);
 144		}
 145	}
 146}
 147
 148static int next_repair_mirror(const struct btrfs_failed_bio *fbio, int cur_mirror)
 149{
 150	if (cur_mirror == fbio->num_copies)
 151		return cur_mirror + 1 - fbio->num_copies;
 152	return cur_mirror + 1;
 153}
 154
 155static int prev_repair_mirror(const struct btrfs_failed_bio *fbio, int cur_mirror)
 156{
 157	if (cur_mirror == 1)
 158		return fbio->num_copies;
 159	return cur_mirror - 1;
 160}
 161
 162static void btrfs_repair_done(struct btrfs_failed_bio *fbio)
 163{
 164	if (atomic_dec_and_test(&fbio->repair_count)) {
 165		btrfs_bio_end_io(fbio->bbio, fbio->bbio->bio.bi_status);
 166		mempool_free(fbio, &btrfs_failed_bio_pool);
 167	}
 168}
 169
 170static void btrfs_end_repair_bio(struct btrfs_bio *repair_bbio,
 171				 struct btrfs_device *dev)
 172{
 173	struct btrfs_failed_bio *fbio = repair_bbio->private;
 174	struct btrfs_inode *inode = repair_bbio->inode;
 175	struct btrfs_fs_info *fs_info = inode->root->fs_info;
 176	/*
 177	 * We can not move forward the saved_iter, as it will be later
 178	 * utilized by repair_bbio again.
 179	 */
 180	struct bvec_iter saved_iter = repair_bbio->saved_iter;
 181	const u32 step = min(fs_info->sectorsize, PAGE_SIZE);
 182	const u64 logical = repair_bbio->saved_iter.bi_sector << SECTOR_SHIFT;
 183	const u32 nr_steps = repair_bbio->saved_iter.bi_size / step;
 184	int mirror = repair_bbio->mirror_num;
 185	phys_addr_t paddrs[BTRFS_MAX_BLOCKSIZE / PAGE_SIZE];
 186	phys_addr_t paddr;
 187	unsigned int slot = 0;
 188
 189	/* Repair bbio should be eaxctly one block sized. */
 190	ASSERT(repair_bbio->saved_iter.bi_size == fs_info->sectorsize);
 191
 192	btrfs_bio_for_each_block(paddr, &repair_bbio->bio, &saved_iter, step) {
 193		ASSERT(slot < nr_steps);
 194		paddrs[slot] = paddr;
 195		slot++;
 196	}
 197
 198	if (repair_bbio->bio.bi_status ||
 199	    !btrfs_data_csum_ok(repair_bbio, dev, 0, paddrs)) {
 200		bio_reset(&repair_bbio->bio, NULL, REQ_OP_READ);
 201		repair_bbio->bio.bi_iter = repair_bbio->saved_iter;
 202
 203		mirror = next_repair_mirror(fbio, mirror);
 204		if (mirror == fbio->bbio->mirror_num) {
 205			btrfs_debug(fs_info, "no mirror left");
 206			fbio->bbio->bio.bi_status = BLK_STS_IOERR;
 207			goto done;
 208		}
 209
 210		btrfs_submit_bbio(repair_bbio, mirror);
 211		return;
 212	}
 213
 214	do {
 215		mirror = prev_repair_mirror(fbio, mirror);
 216		btrfs_repair_io_failure(fs_info, btrfs_ino(inode),
 217				  repair_bbio->file_offset, fs_info->sectorsize,
 218				  logical, paddrs, step, mirror);
 219	} while (mirror != fbio->bbio->mirror_num);
 220
 221done:
 222	btrfs_repair_done(fbio);
 223	bio_put(&repair_bbio->bio);
 224}
 225
 226/*
 227 * Try to kick off a repair read to the next available mirror for a bad sector.
 228 *
 229 * This primarily tries to recover good data to serve the actual read request,
 230 * but also tries to write the good data back to the bad mirror(s) when a
 231 * read succeeded to restore the redundancy.
 232 */
 233static struct btrfs_failed_bio *repair_one_sector(struct btrfs_bio *failed_bbio,
 234						  u32 bio_offset,
 235						  phys_addr_t paddrs[],
 236						  struct btrfs_failed_bio *fbio)
 237{
 238	struct btrfs_inode *inode = failed_bbio->inode;
 239	struct btrfs_fs_info *fs_info = inode->root->fs_info;
 240	const u32 sectorsize = fs_info->sectorsize;
 241	const u32 step = min(fs_info->sectorsize, PAGE_SIZE);
 242	const u32 nr_steps = sectorsize / step;
 243	/*
 244	 * For bs > ps cases, the saved_iter can be partially moved forward.
 245	 * In that case we should round it down to the block boundary.
 246	 */
 247	const u64 logical = round_down(failed_bbio->saved_iter.bi_sector << SECTOR_SHIFT,
 248				       sectorsize);
 249	struct btrfs_bio *repair_bbio;
 250	struct bio *repair_bio;
 251	int num_copies;
 252	int mirror;
 253
 254	btrfs_debug(fs_info, "repair read error: read error at %llu",
 255		    failed_bbio->file_offset + bio_offset);
 256
 257	num_copies = btrfs_num_copies(fs_info, logical, sectorsize);
 258	if (num_copies == 1) {
 259		btrfs_debug(fs_info, "no copy to repair from");
 260		failed_bbio->bio.bi_status = BLK_STS_IOERR;
 261		return fbio;
 262	}
 263
 264	if (!fbio) {
 265		fbio = mempool_alloc(&btrfs_failed_bio_pool, GFP_NOFS);
 266		fbio->bbio = failed_bbio;
 267		fbio->num_copies = num_copies;
 268		atomic_set(&fbio->repair_count, 1);
 269	}
 270
 271	atomic_inc(&fbio->repair_count);
 272
 273	repair_bio = bio_alloc_bioset(NULL, nr_steps, REQ_OP_READ, GFP_NOFS,
 274				      &btrfs_repair_bioset);
 275	repair_bio->bi_iter.bi_sector = logical >> SECTOR_SHIFT;
 276	for (int i = 0; i < nr_steps; i++) {
 277		int ret;
 278
 279		ASSERT(offset_in_page(paddrs[i]) + step <= PAGE_SIZE);
 280
 281		ret = bio_add_page(repair_bio, phys_to_page(paddrs[i]), step,
 282				   offset_in_page(paddrs[i]));
 283		ASSERT(ret == step);
 284	}
 285
 286	repair_bbio = btrfs_bio(repair_bio);
 287	btrfs_bio_init(repair_bbio, failed_bbio->inode, failed_bbio->file_offset + bio_offset,
 288		       NULL, fbio);
 289
 290	mirror = next_repair_mirror(fbio, failed_bbio->mirror_num);
 291	btrfs_debug(fs_info, "submitting repair read to mirror %d", mirror);
 292	btrfs_submit_bbio(repair_bbio, mirror);
 293	return fbio;
 294}
 295
 296static void btrfs_check_read_bio(struct btrfs_bio *bbio, struct btrfs_device *dev)
 297{
 298	struct btrfs_inode *inode = bbio->inode;
 299	struct btrfs_fs_info *fs_info = inode->root->fs_info;
 300	const u32 sectorsize = fs_info->sectorsize;
 301	const u32 step = min(sectorsize, PAGE_SIZE);
 302	const u32 nr_steps = sectorsize / step;
 303	struct bvec_iter *iter = &bbio->saved_iter;
 304	blk_status_t status = bbio->bio.bi_status;
 305	struct btrfs_failed_bio *fbio = NULL;
 306	phys_addr_t paddrs[BTRFS_MAX_BLOCKSIZE / PAGE_SIZE];
 307	phys_addr_t paddr;
 308	u32 offset = 0;
 309
 310	/* Read-repair requires the inode field to be set by the submitter. */
 311	ASSERT(inode);
 312
 313	/*
 314	 * Hand off repair bios to the repair code as there is no upper level
 315	 * submitter for them.
 316	 */
 317	if (bbio->bio.bi_pool == &btrfs_repair_bioset) {
 318		btrfs_end_repair_bio(bbio, dev);
 319		return;
 320	}
 321
 322	/* Clear the I/O error. A failed repair will reset it. */
 323	bbio->bio.bi_status = BLK_STS_OK;
 324
 325	btrfs_bio_for_each_block(paddr, &bbio->bio, iter, step) {
 326		paddrs[(offset / step) % nr_steps] = paddr;
 327		offset += step;
 328
 329		if (IS_ALIGNED(offset, sectorsize)) {
 330			if (status ||
 331			    !btrfs_data_csum_ok(bbio, dev, offset - sectorsize, paddrs))
 332				fbio = repair_one_sector(bbio, offset - sectorsize,
 333							 paddrs, fbio);
 334		}
 335	}
 336	if (bbio->csum != bbio->csum_inline)
 337		kvfree(bbio->csum);
 338
 339	if (fbio)
 340		btrfs_repair_done(fbio);
 341	else
 342		btrfs_bio_end_io(bbio, bbio->bio.bi_status);
 343}
 344
 345static void btrfs_log_dev_io_error(const struct bio *bio, struct btrfs_device *dev)
 346{
 347	if (!dev || !dev->bdev)
 348		return;
 349	if (bio->bi_status != BLK_STS_IOERR && bio->bi_status != BLK_STS_TARGET)
 350		return;
 351
 352	if (btrfs_op(bio) == BTRFS_MAP_WRITE)
 353		btrfs_dev_stat_inc_and_print(dev, BTRFS_DEV_STAT_WRITE_ERRS);
 354	else if (!(bio->bi_opf & REQ_RAHEAD))
 355		btrfs_dev_stat_inc_and_print(dev, BTRFS_DEV_STAT_READ_ERRS);
 356	if (bio->bi_opf & REQ_PREFLUSH)
 357		btrfs_dev_stat_inc_and_print(dev, BTRFS_DEV_STAT_FLUSH_ERRS);
 358}
 359
 360static struct workqueue_struct *btrfs_end_io_wq(const struct btrfs_fs_info *fs_info,
 361						const struct bio *bio)
 362{
 363	if (bio->bi_opf & REQ_META)
 364		return fs_info->endio_meta_workers;
 365	return fs_info->endio_workers;
 366}
 367
 368static void simple_end_io_work(struct work_struct *work)
 369{
 370	struct btrfs_bio *bbio = container_of(work, struct btrfs_bio, end_io_work);
 371	struct bio *bio = &bbio->bio;
 372
 373	if (bio_op(bio) == REQ_OP_READ) {
 374		/* Metadata reads are checked and repaired by the submitter. */
 375		if (is_data_bbio(bbio))
 376			return btrfs_check_read_bio(bbio, bbio->bio.bi_private);
 377		return btrfs_bio_end_io(bbio, bbio->bio.bi_status);
 378	}
 379	if (bio_is_zone_append(bio) && !bio->bi_status)
 380		btrfs_record_physical_zoned(bbio);
 381	btrfs_bio_end_io(bbio, bbio->bio.bi_status);
 382}
 383
 384static void btrfs_simple_end_io(struct bio *bio)
 385{
 386	struct btrfs_bio *bbio = btrfs_bio(bio);
 387	struct btrfs_device *dev = bio->bi_private;
 388	struct btrfs_fs_info *fs_info = bbio->inode->root->fs_info;
 389
 390	btrfs_bio_counter_dec(fs_info);
 391
 392	if (bio->bi_status)
 393		btrfs_log_dev_io_error(bio, dev);
 394
 395	INIT_WORK(&bbio->end_io_work, simple_end_io_work);
 396	queue_work(btrfs_end_io_wq(fs_info, bio), &bbio->end_io_work);
 397}
 398
 399static void btrfs_raid56_end_io(struct bio *bio)
 400{
 401	struct btrfs_io_context *bioc = bio->bi_private;
 402	struct btrfs_bio *bbio = btrfs_bio(bio);
 403
 404	/* RAID56 endio is always handled in workqueue. */
 405	ASSERT(in_task());
 406
 407	btrfs_bio_counter_dec(bioc->fs_info);
 408	bbio->mirror_num = bioc->mirror_num;
 409	if (bio_op(bio) == REQ_OP_READ && is_data_bbio(bbio))
 410		btrfs_check_read_bio(bbio, NULL);
 411	else
 412		btrfs_bio_end_io(bbio, bbio->bio.bi_status);
 413
 414	btrfs_put_bioc(bioc);
 415}
 416
 417static void orig_write_end_io_work(struct work_struct *work)
 418{
 419	struct btrfs_bio *bbio = container_of(work, struct btrfs_bio, end_io_work);
 420	struct bio *bio = &bbio->bio;
 421	struct btrfs_io_stripe *stripe = bio->bi_private;
 422	struct btrfs_io_context *bioc = stripe->bioc;
 423
 424	btrfs_bio_counter_dec(bioc->fs_info);
 425
 426	if (bio->bi_status) {
 427		atomic_inc(&bioc->error);
 428		btrfs_log_dev_io_error(bio, stripe->dev);
 429	}
 430
 431	/*
 432	 * Only send an error to the higher layers if it is beyond the tolerance
 433	 * threshold.
 434	 */
 435	if (atomic_read(&bioc->error) > bioc->max_errors)
 436		bio->bi_status = BLK_STS_IOERR;
 437	else
 438		bio->bi_status = BLK_STS_OK;
 439
 440	if (bio_is_zone_append(bio) && !bio->bi_status)
 441		stripe->physical = bio->bi_iter.bi_sector << SECTOR_SHIFT;
 442
 443	btrfs_bio_end_io(bbio, bbio->bio.bi_status);
 444	btrfs_put_bioc(bioc);
 445}
 446
 447static void btrfs_orig_write_end_io(struct bio *bio)
 448{
 449	struct btrfs_bio *bbio = btrfs_bio(bio);
 450
 451	INIT_WORK(&bbio->end_io_work, orig_write_end_io_work);
 452	queue_work(btrfs_end_io_wq(bbio->inode->root->fs_info, bio), &bbio->end_io_work);
 453}
 454
 455static void clone_write_end_io_work(struct work_struct *work)
 456{
 457	struct btrfs_bio *bbio = container_of(work, struct btrfs_bio, end_io_work);
 458	struct bio *bio = &bbio->bio;
 459	struct btrfs_io_stripe *stripe = bio->bi_private;
 460
 461	if (bio->bi_status) {
 462		atomic_inc(&stripe->bioc->error);
 463		btrfs_log_dev_io_error(bio, stripe->dev);
 464	} else if (bio_is_zone_append(bio)) {
 465		stripe->physical = bio->bi_iter.bi_sector << SECTOR_SHIFT;
 466	}
 467
 468	/* Pass on control to the original bio this one was cloned from */
 469	bio_endio(stripe->bioc->orig_bio);
 470	bio_put(bio);
 471}
 472
 473static void btrfs_clone_write_end_io(struct bio *bio)
 474{
 475	struct btrfs_bio *bbio = btrfs_bio(bio);
 476
 477	INIT_WORK(&bbio->end_io_work, clone_write_end_io_work);
 478	queue_work(btrfs_end_io_wq(bbio->inode->root->fs_info, bio), &bbio->end_io_work);
 479}
 480
 481static void btrfs_submit_dev_bio(struct btrfs_device *dev, struct bio *bio)
 482{
 483	u64 physical = bio->bi_iter.bi_sector << SECTOR_SHIFT;
 484
 485	if (!dev || !dev->bdev ||
 486	    test_bit(BTRFS_DEV_STATE_MISSING, &dev->dev_state) ||
 487	    (btrfs_op(bio) == BTRFS_MAP_WRITE &&
 488	     !test_bit(BTRFS_DEV_STATE_WRITEABLE, &dev->dev_state))) {
 489		bio_io_error(bio);
 490		return;
 491	}
 492
 493	bio_set_dev(bio, dev->bdev);
 494
 495	/*
 496	 * For zone append writing, bi_sector must point the beginning of the
 497	 * zone
 498	 */
 499	if (btrfs_bio(bio)->can_use_append && btrfs_dev_is_sequential(dev, physical)) {
 500		u64 zone_start = round_down(physical, dev->fs_info->zone_size);
 501
 502		ASSERT(btrfs_dev_is_sequential(dev, physical));
 503		bio->bi_iter.bi_sector = zone_start >> SECTOR_SHIFT;
 504		bio->bi_opf &= ~REQ_OP_WRITE;
 505		bio->bi_opf |= REQ_OP_ZONE_APPEND;
 506	}
 507	btrfs_debug(dev->fs_info,
 508	"%s: rw %d 0x%x, sector=%llu, dev=%lu (%s id %llu), size=%u",
 509		__func__, bio_op(bio), bio->bi_opf, bio->bi_iter.bi_sector,
 510		(unsigned long)dev->bdev->bd_dev, btrfs_dev_name(dev),
 511		dev->devid, bio->bi_iter.bi_size);
 512
 513	/*
 514	 * Track reads if tracking is enabled; ignore I/O operations before the
 515	 * filesystem is fully initialized.
 516	 */
 517	if (dev->fs_devices->collect_fs_stats && bio_op(bio) == REQ_OP_READ && dev->fs_info)
 518		percpu_counter_add(&dev->fs_info->stats_read_blocks,
 519				   bio->bi_iter.bi_size >> dev->fs_info->sectorsize_bits);
 520
 521	if (bio->bi_opf & REQ_BTRFS_CGROUP_PUNT)
 522		blkcg_punt_bio_submit(bio);
 523	else
 524		submit_bio(bio);
 525}
 526
 527static void btrfs_submit_mirrored_bio(struct btrfs_io_context *bioc, int dev_nr)
 528{
 529	struct bio *orig_bio = bioc->orig_bio, *bio;
 530	struct btrfs_bio *orig_bbio = btrfs_bio(orig_bio);
 531
 532	ASSERT(bio_op(orig_bio) != REQ_OP_READ);
 533
 534	/* Reuse the bio embedded into the btrfs_bio for the last mirror */
 535	if (dev_nr == bioc->num_stripes - 1) {
 536		bio = orig_bio;
 537		bio->bi_end_io = btrfs_orig_write_end_io;
 538	} else {
 539		/* We need to use endio_work to run end_io in task context. */
 540		bio = bio_alloc_clone(NULL, orig_bio, GFP_NOFS, &btrfs_bioset);
 541		bio_inc_remaining(orig_bio);
 542		btrfs_bio_init(btrfs_bio(bio), orig_bbio->inode,
 543			       orig_bbio->file_offset, NULL, NULL);
 544		bio->bi_end_io = btrfs_clone_write_end_io;
 545	}
 546
 547	bio->bi_private = &bioc->stripes[dev_nr];
 548	bio->bi_iter.bi_sector = bioc->stripes[dev_nr].physical >> SECTOR_SHIFT;
 549	bioc->stripes[dev_nr].bioc = bioc;
 550	bioc->size = bio->bi_iter.bi_size;
 551	btrfs_submit_dev_bio(bioc->stripes[dev_nr].dev, bio);
 552}
 553
 554static void btrfs_submit_bio(struct bio *bio, struct btrfs_io_context *bioc,
 555			     struct btrfs_io_stripe *smap, int mirror_num)
 556{
 557	if (!bioc) {
 558		/* Single mirror read/write fast path. */
 559		btrfs_bio(bio)->mirror_num = mirror_num;
 560		bio->bi_iter.bi_sector = smap->physical >> SECTOR_SHIFT;
 561		if (bio_op(bio) != REQ_OP_READ)
 562			btrfs_bio(bio)->orig_physical = smap->physical;
 563		bio->bi_private = smap->dev;
 564		bio->bi_end_io = btrfs_simple_end_io;
 565		btrfs_submit_dev_bio(smap->dev, bio);
 566	} else if (bioc->map_type & BTRFS_BLOCK_GROUP_RAID56_MASK) {
 567		/* Parity RAID write or read recovery. */
 568		bio->bi_private = bioc;
 569		bio->bi_end_io = btrfs_raid56_end_io;
 570		if (bio_op(bio) == REQ_OP_READ)
 571			raid56_parity_recover(bio, bioc, mirror_num);
 572		else
 573			raid56_parity_write(bio, bioc);
 574	} else {
 575		/* Write to multiple mirrors. */
 576		int total_devs = bioc->num_stripes;
 577
 578		bioc->orig_bio = bio;
 579		for (int dev_nr = 0; dev_nr < total_devs; dev_nr++)
 580			btrfs_submit_mirrored_bio(bioc, dev_nr);
 581	}
 582}
 583
 584static int btrfs_bio_csum(struct btrfs_bio *bbio)
 585{
 586	if (bbio->bio.bi_opf & REQ_META)
 587		return btree_csum_one_bio(bbio);
 588#ifdef CONFIG_BTRFS_EXPERIMENTAL
 589	return btrfs_csum_one_bio(bbio, true);
 590#else
 591	return btrfs_csum_one_bio(bbio, false);
 592#endif
 593}
 594
 595/*
 596 * Async submit bios are used to offload expensive checksumming onto the worker
 597 * threads.
 598 */
 599struct async_submit_bio {
 600	struct btrfs_bio *bbio;
 601	struct btrfs_io_context *bioc;
 602	struct btrfs_io_stripe smap;
 603	int mirror_num;
 604	struct btrfs_work work;
 605};
 606
 607/*
 608 * In order to insert checksums into the metadata in large chunks, we wait
 609 * until bio submission time.   All the pages in the bio are checksummed and
 610 * sums are attached onto the ordered extent record.
 611 *
 612 * At IO completion time the csums attached on the ordered extent record are
 613 * inserted into the btree.
 614 */
 615static void run_one_async_start(struct btrfs_work *work)
 616{
 617	struct async_submit_bio *async =
 618		container_of(work, struct async_submit_bio, work);
 619	int ret;
 620
 621	ret = btrfs_bio_csum(async->bbio);
 622	if (ret)
 623		async->bbio->bio.bi_status = errno_to_blk_status(ret);
 624}
 625
 626/*
 627 * In order to insert checksums into the metadata in large chunks, we wait
 628 * until bio submission time.   All the pages in the bio are checksummed and
 629 * sums are attached onto the ordered extent record.
 630 *
 631 * At IO completion time the csums attached on the ordered extent record are
 632 * inserted into the tree.
 633 *
 634 * If called with @do_free == true, then it will free the work struct.
 635 */
 636static void run_one_async_done(struct btrfs_work *work, bool do_free)
 637{
 638	struct async_submit_bio *async =
 639		container_of(work, struct async_submit_bio, work);
 640	struct bio *bio = &async->bbio->bio;
 641
 642	if (do_free) {
 643		kfree(container_of(work, struct async_submit_bio, work));
 644		return;
 645	}
 646
 647	/* If an error occurred we just want to clean up the bio and move on. */
 648	if (bio->bi_status) {
 649		btrfs_bio_end_io(async->bbio, bio->bi_status);
 650		return;
 651	}
 652
 653	/*
 654	 * All of the bios that pass through here are from async helpers.
 655	 * Use REQ_BTRFS_CGROUP_PUNT to issue them from the owning cgroup's
 656	 * context.  This changes nothing when cgroups aren't in use.
 657	 */
 658	bio->bi_opf |= REQ_BTRFS_CGROUP_PUNT;
 659	btrfs_submit_bio(bio, async->bioc, &async->smap, async->mirror_num);
 660}
 661
 662static bool should_async_write(struct btrfs_bio *bbio)
 663{
 664	struct btrfs_fs_info *fs_info = bbio->inode->root->fs_info;
 665	bool auto_csum_mode = true;
 666
 667#ifdef CONFIG_BTRFS_EXPERIMENTAL
 668	struct btrfs_fs_devices *fs_devices = fs_info->fs_devices;
 669	enum btrfs_offload_csum_mode csum_mode = READ_ONCE(fs_devices->offload_csum_mode);
 670
 671	if (csum_mode == BTRFS_OFFLOAD_CSUM_FORCE_ON)
 672		return true;
 673	/*
 674	 * Write bios will calculate checksum and submit bio at the same time.
 675	 * Unless explicitly required don't offload serial csum calculate and bio
 676	 * submit into a workqueue.
 677	 */
 678	return false;
 679#endif
 680
 681	/* Submit synchronously if the checksum implementation is fast. */
 682	if (auto_csum_mode && test_bit(BTRFS_FS_CSUM_IMPL_FAST, &fs_info->flags))
 683		return false;
 684
 685	/*
 686	 * Try to defer the submission to a workqueue to parallelize the
 687	 * checksum calculation unless the I/O is issued synchronously.
 688	 */
 689	if (op_is_sync(bbio->bio.bi_opf))
 690		return false;
 691
 692	/* Zoned devices require I/O to be submitted in order. */
 693	if ((bbio->bio.bi_opf & REQ_META) && btrfs_is_zoned(fs_info))
 694		return false;
 695
 696	return true;
 697}
 698
 699/*
 700 * Submit bio to an async queue.
 701 *
 702 * Return true if the work has been successfully submitted, else false.
 703 */
 704static bool btrfs_wq_submit_bio(struct btrfs_bio *bbio,
 705				struct btrfs_io_context *bioc,
 706				struct btrfs_io_stripe *smap, int mirror_num)
 707{
 708	struct btrfs_fs_info *fs_info = bbio->inode->root->fs_info;
 709	struct async_submit_bio *async;
 710
 711	async = kmalloc(sizeof(*async), GFP_NOFS);
 712	if (!async)
 713		return false;
 714
 715	async->bbio = bbio;
 716	async->bioc = bioc;
 717	async->smap = *smap;
 718	async->mirror_num = mirror_num;
 719
 720	btrfs_init_work(&async->work, run_one_async_start, run_one_async_done);
 721	btrfs_queue_work(fs_info->workers, &async->work);
 722	return true;
 723}
 724
 725static u64 btrfs_append_map_length(struct btrfs_bio *bbio, u64 map_length)
 726{
 727	struct btrfs_fs_info *fs_info = bbio->inode->root->fs_info;
 728	unsigned int nr_segs;
 729	int sector_offset;
 730
 731	map_length = min(map_length, fs_info->max_zone_append_size);
 732	sector_offset = bio_split_rw_at(&bbio->bio, &fs_info->limits,
 733					&nr_segs, map_length);
 734	if (sector_offset) {
 735		/*
 736		 * bio_split_rw_at() could split at a size smaller than our
 737		 * sectorsize and thus cause unaligned I/Os.  Fix that by
 738		 * always rounding down to the nearest boundary.
 739		 */
 740		return ALIGN_DOWN(sector_offset << SECTOR_SHIFT, fs_info->sectorsize);
 741	}
 742	return map_length;
 743}
 744
 745static bool btrfs_submit_chunk(struct btrfs_bio *bbio, int mirror_num)
 746{
 747	struct btrfs_inode *inode = bbio->inode;
 748	struct btrfs_fs_info *fs_info = inode->root->fs_info;
 749	struct bio *bio = &bbio->bio;
 750	u64 logical = bio->bi_iter.bi_sector << SECTOR_SHIFT;
 751	u64 length = bio->bi_iter.bi_size;
 752	u64 map_length = length;
 753	struct btrfs_io_context *bioc = NULL;
 754	struct btrfs_io_stripe smap;
 755	blk_status_t status;
 756	int ret;
 757
 758	if (bbio->is_scrub || btrfs_is_data_reloc_root(inode->root))
 759		smap.rst_search_commit_root = true;
 760	else
 761		smap.rst_search_commit_root = false;
 762
 763	btrfs_bio_counter_inc_blocked(fs_info);
 764	ret = btrfs_map_block(fs_info, btrfs_op(bio), logical, &map_length,
 765			      &bioc, &smap, &mirror_num);
 766	if (ret) {
 767		status = errno_to_blk_status(ret);
 768		btrfs_bio_counter_dec(fs_info);
 769		goto end_bbio;
 770	}
 771
 772	/*
 773	 * For fscrypt writes we will get the encrypted bio after we've remapped
 774	 * our bio to the physical disk location, so we need to save the
 775	 * original bytenr so we know what we're checksumming.
 776	 */
 777	if (bio_op(bio) == REQ_OP_WRITE && is_data_bbio(bbio))
 778		bbio->orig_logical = logical;
 779
 780	bbio->can_use_append = btrfs_use_zone_append(bbio);
 781
 782	map_length = min(map_length, length);
 783	if (bbio->can_use_append)
 784		map_length = btrfs_append_map_length(bbio, map_length);
 785
 786	if (map_length < length) {
 787		struct btrfs_bio *split;
 788
 789		split = btrfs_split_bio(fs_info, bbio, map_length);
 790		if (IS_ERR(split)) {
 791			status = errno_to_blk_status(PTR_ERR(split));
 792			btrfs_bio_counter_dec(fs_info);
 793			goto end_bbio;
 794		}
 795		bbio = split;
 796		bio = &bbio->bio;
 797	}
 798
 799	/*
 800	 * Save the iter for the end_io handler and preload the checksums for
 801	 * data reads.
 802	 */
 803	if (bio_op(bio) == REQ_OP_READ && is_data_bbio(bbio)) {
 804		bbio->saved_iter = bio->bi_iter;
 805		ret = btrfs_lookup_bio_sums(bbio);
 806		status = errno_to_blk_status(ret);
 807		if (status)
 808			goto fail;
 809	}
 810
 811	if (btrfs_op(bio) == BTRFS_MAP_WRITE) {
 812		if (is_data_bbio(bbio) && bioc && bioc->use_rst) {
 813			/*
 814			 * No locking for the list update, as we only add to
 815			 * the list in the I/O submission path, and list
 816			 * iteration only happens in the completion path, which
 817			 * can't happen until after the last submission.
 818			 */
 819			btrfs_get_bioc(bioc);
 820			list_add_tail(&bioc->rst_ordered_entry, &bbio->ordered->bioc_list);
 821		}
 822
 823		/*
 824		 * Csum items for reloc roots have already been cloned at this
 825		 * point, so they are handled as part of the no-checksum case.
 826		 */
 827		if (!(inode->flags & BTRFS_INODE_NODATASUM) &&
 828		    !test_bit(BTRFS_FS_STATE_NO_DATA_CSUMS, &fs_info->fs_state) &&
 829		    !btrfs_is_data_reloc_root(inode->root)) {
 830			if (should_async_write(bbio) &&
 831			    btrfs_wq_submit_bio(bbio, bioc, &smap, mirror_num))
 832				goto done;
 833
 834			ret = btrfs_bio_csum(bbio);
 835			status = errno_to_blk_status(ret);
 836			if (status)
 837				goto fail;
 838		} else if (bbio->can_use_append ||
 839			   (btrfs_is_zoned(fs_info) && inode &&
 840			    inode->flags & BTRFS_INODE_NODATASUM)) {
 841			ret = btrfs_alloc_dummy_sum(bbio);
 842			status = errno_to_blk_status(ret);
 843			if (status)
 844				goto fail;
 845		}
 846	}
 847
 848	btrfs_submit_bio(bio, bioc, &smap, mirror_num);
 849done:
 850	return map_length == length;
 851
 852fail:
 853	btrfs_bio_counter_dec(fs_info);
 854	/*
 855	 * We have split the original bbio, now we have to end both the current
 856	 * @bbio and remaining one, as the remaining one will never be submitted.
 857	 */
 858	if (map_length < length) {
 859		struct btrfs_bio *remaining = bbio->private;
 860
 861		ASSERT(bbio->bio.bi_pool == &btrfs_clone_bioset);
 862		ASSERT(remaining);
 863
 864		btrfs_bio_end_io(remaining, status);
 865	}
 866end_bbio:
 867	btrfs_bio_end_io(bbio, status);
 868	/* Do not submit another chunk */
 869	return true;
 870}
 871
 872static void assert_bbio_alignment(struct btrfs_bio *bbio)
 873{
 874#ifdef CONFIG_BTRFS_ASSERT
 875	struct btrfs_fs_info *fs_info = bbio->inode->root->fs_info;
 876	struct bio_vec bvec;
 877	struct bvec_iter iter;
 878	const u32 blocksize = fs_info->sectorsize;
 879	const u32 alignment = min(blocksize, PAGE_SIZE);
 880	const u64 logical = bbio->bio.bi_iter.bi_sector << SECTOR_SHIFT;
 881	const u32 length = bbio->bio.bi_iter.bi_size;
 882
 883	/* The logical and length should still be aligned to blocksize. */
 884	ASSERT(IS_ALIGNED(logical, blocksize) && IS_ALIGNED(length, blocksize) &&
 885	       length != 0, "root=%llu inode=%llu logical=%llu length=%u",
 886	       btrfs_root_id(bbio->inode->root),
 887	       btrfs_ino(bbio->inode), logical, length);
 888
 889	bio_for_each_bvec(bvec, &bbio->bio, iter)
 890		ASSERT(IS_ALIGNED(bvec.bv_offset, alignment) &&
 891		       IS_ALIGNED(bvec.bv_len, alignment),
 892		"root=%llu inode=%llu logical=%llu length=%u index=%u bv_offset=%u bv_len=%u",
 893		btrfs_root_id(bbio->inode->root),
 894		btrfs_ino(bbio->inode), logical, length, iter.bi_idx,
 895		bvec.bv_offset, bvec.bv_len);
 896#endif
 897}
 898
 899void btrfs_submit_bbio(struct btrfs_bio *bbio, int mirror_num)
 900{
 901	/* If bbio->inode is not populated, its file_offset must be 0. */
 902	ASSERT(bbio->inode || bbio->file_offset == 0);
 903
 904	assert_bbio_alignment(bbio);
 905
 906	while (!btrfs_submit_chunk(bbio, mirror_num))
 907		;
 908}
 909
 910/*
 911 * Submit a repair write.
 912 *
 913 * This bypasses btrfs_submit_bbio() deliberately, as that writes all copies in a
 914 * RAID setup.  Here we only want to write the one bad copy, so we do the
 915 * mapping ourselves and submit the bio directly.
 916 *
 917 * The I/O is issued synchronously to block the repair read completion from
 918 * freeing the bio.
 919 *
 920 * @ino:	Offending inode number
 921 * @fileoff:	File offset inside the inode
 922 * @length:	Length of the repair write
 923 * @logical:	Logical address of the range
 924 * @paddrs:	Physical address array of the content
 925 * @step:	Length of for each paddrs
 926 * @mirror_num: Mirror number to write to. Must not be zero
 927 */
 928int btrfs_repair_io_failure(struct btrfs_fs_info *fs_info, u64 ino, u64 fileoff,
 929			    u32 length, u64 logical, const phys_addr_t paddrs[],
 930			    unsigned int step, int mirror_num)
 931{
 932	const u32 nr_steps = DIV_ROUND_UP_POW2(length, step);
 933	struct btrfs_io_stripe smap = { 0 };
 934	struct bio *bio = NULL;
 935	int ret = 0;
 936
 937	BUG_ON(!mirror_num);
 938
 939	/* Basic alignment checks. */
 940	ASSERT(IS_ALIGNED(logical, fs_info->sectorsize));
 941	ASSERT(IS_ALIGNED(length, fs_info->sectorsize));
 942	ASSERT(IS_ALIGNED(fileoff, fs_info->sectorsize));
 943	/* Either it's a single data or metadata block. */
 944	ASSERT(length <= BTRFS_MAX_BLOCKSIZE);
 945	ASSERT(step <= length);
 946	ASSERT(is_power_of_2(step));
 947
 948	/*
 949	 * The fs either mounted RO or hit critical errors, no need
 950	 * to continue repairing.
 951	 */
 952	if (unlikely(sb_rdonly(fs_info->sb)))
 953		return 0;
 954
 955	if (btrfs_repair_one_zone(fs_info, logical))
 956		return 0;
 957
 958	/*
 959	 * Avoid races with device replace and make sure our bioc has devices
 960	 * associated to its stripes that don't go away while we are doing the
 961	 * read repair operation.
 962	 */
 963	btrfs_bio_counter_inc_blocked(fs_info);
 964	ret = btrfs_map_repair_block(fs_info, &smap, logical, length, mirror_num);
 965	if (ret < 0)
 966		goto out_counter_dec;
 967
 968	if (unlikely(!smap.dev->bdev ||
 969		     !test_bit(BTRFS_DEV_STATE_WRITEABLE, &smap.dev->dev_state))) {
 970		ret = -EIO;
 971		goto out_counter_dec;
 972	}
 973
 974	bio = bio_alloc(smap.dev->bdev, nr_steps, REQ_OP_WRITE | REQ_SYNC, GFP_NOFS);
 975	bio->bi_iter.bi_sector = smap.physical >> SECTOR_SHIFT;
 976	for (int i = 0; i < nr_steps; i++) {
 977		ret = bio_add_page(bio, phys_to_page(paddrs[i]), step, offset_in_page(paddrs[i]));
 978		/* We should have allocated enough slots to contain all the different pages. */
 979		ASSERT(ret == step);
 980	}
 981	ret = submit_bio_wait(bio);
 982	bio_put(bio);
 983	if (ret) {
 984		/* try to remap that extent elsewhere? */
 985		btrfs_dev_stat_inc_and_print(smap.dev, BTRFS_DEV_STAT_WRITE_ERRS);
 986		goto out_counter_dec;
 987	}
 988
 989	btrfs_info_rl(fs_info,
 990		"read error corrected: ino %llu off %llu (dev %s sector %llu)",
 991			     ino, fileoff, btrfs_dev_name(smap.dev),
 992			     smap.physical >> SECTOR_SHIFT);
 993	ret = 0;
 994
 995out_counter_dec:
 996	btrfs_bio_counter_dec(fs_info);
 997	return ret;
 998}
 999
1000/*
1001 * Submit a btrfs_bio based repair write.
1002 *
1003 * If @dev_replace is true, the write would be submitted to dev-replace target.
1004 */
1005void btrfs_submit_repair_write(struct btrfs_bio *bbio, int mirror_num, bool dev_replace)
1006{
1007	struct btrfs_fs_info *fs_info = bbio->inode->root->fs_info;
1008	u64 logical = bbio->bio.bi_iter.bi_sector << SECTOR_SHIFT;
1009	u64 length = bbio->bio.bi_iter.bi_size;
1010	struct btrfs_io_stripe smap = { 0 };
1011	int ret;
1012
1013	ASSERT(mirror_num > 0);
1014	ASSERT(btrfs_op(&bbio->bio) == BTRFS_MAP_WRITE);
1015	ASSERT(!is_data_inode(bbio->inode));
1016	ASSERT(bbio->is_scrub);
1017
1018	btrfs_bio_counter_inc_blocked(fs_info);
1019	ret = btrfs_map_repair_block(fs_info, &smap, logical, length, mirror_num);
1020	if (ret < 0)
1021		goto fail;
1022
1023	if (dev_replace) {
1024		ASSERT(smap.dev == fs_info->dev_replace.srcdev);
1025		smap.dev = fs_info->dev_replace.tgtdev;
1026	}
1027	btrfs_submit_bio(&bbio->bio, NULL, &smap, mirror_num);
1028	return;
1029
1030fail:
1031	btrfs_bio_counter_dec(fs_info);
1032	btrfs_bio_end_io(bbio, errno_to_blk_status(ret));
1033}
1034
1035int __init btrfs_bioset_init(void)
1036{
1037	if (bioset_init(&btrfs_bioset, BIO_POOL_SIZE,
1038			offsetof(struct btrfs_bio, bio),
1039			BIOSET_NEED_BVECS))
1040		return -ENOMEM;
1041	if (bioset_init(&btrfs_clone_bioset, BIO_POOL_SIZE,
1042			offsetof(struct btrfs_bio, bio), 0))
1043		goto out;
1044	if (bioset_init(&btrfs_repair_bioset, BIO_POOL_SIZE,
1045			offsetof(struct btrfs_bio, bio),
1046			BIOSET_NEED_BVECS))
1047		goto out;
1048	if (mempool_init_kmalloc_pool(&btrfs_failed_bio_pool, BIO_POOL_SIZE,
1049				      sizeof(struct btrfs_failed_bio)))
1050		goto out;
1051	return 0;
1052
1053out:
1054	btrfs_bioset_exit();
1055	return -ENOMEM;
1056}
1057
1058void __cold btrfs_bioset_exit(void)
1059{
1060	mempool_exit(&btrfs_failed_bio_pool);
1061	bioset_exit(&btrfs_repair_bioset);
1062	bioset_exit(&btrfs_clone_bioset);
1063	bioset_exit(&btrfs_bioset);
1064}