1 // SPDX-License-Identifier: GPL-2.0
3 * Some low level IO code, and hacks for various block layer limitations
5 * Copyright 2010, 2011 Kent Overstreet <kent.overstreet@gmail.com>
6 * Copyright 2012 Google, Inc.
10 #include "alloc_background.h"
11 #include "alloc_foreground.h"
12 #include "btree_update.h"
17 #include "data_update.h"
18 #include "disk_groups.h"
24 #include "subvolume.h"
27 #include <linux/sched/mm.h>
29 #ifndef CONFIG_BCACHEFS_NO_LATENCY_ACCT
31 static bool bch2_target_congested(struct bch_fs *c, u16 target)
33 const struct bch_devs_mask *devs;
34 unsigned d, nr = 0, total = 0;
35 u64 now = local_clock(), last;
43 devs = bch2_target_to_mask(c, target) ?:
44 &c->rw_devs[BCH_DATA_user];
46 for_each_set_bit(d, devs->d, BCH_SB_MEMBERS_MAX) {
47 ca = rcu_dereference(c->devs[d]);
51 congested = atomic_read(&ca->congested);
52 last = READ_ONCE(ca->congested_last);
53 if (time_after64(now, last))
54 congested -= (now - last) >> 12;
56 total += max(congested, 0LL);
61 return bch2_rand_range(nr * CONGESTED_MAX) < total;
66 static bool bch2_target_congested(struct bch_fs *c, u16 target)
73 /* Cache promotion on read */
79 struct rhash_head hash;
82 struct data_update write;
83 struct bio_vec bi_inline_vecs[]; /* must be last */
86 static const struct rhashtable_params bch_promote_params = {
87 .head_offset = offsetof(struct promote_op, hash),
88 .key_offset = offsetof(struct promote_op, pos),
89 .key_len = sizeof(struct bpos),
92 static inline int should_promote(struct bch_fs *c, struct bkey_s_c k,
94 struct bch_io_opts opts,
97 BUG_ON(!opts.promote_target);
99 if (!(flags & BCH_READ_MAY_PROMOTE))
100 return -BCH_ERR_nopromote_may_not;
102 if (bch2_bkey_has_target(c, k, opts.promote_target))
103 return -BCH_ERR_nopromote_already_promoted;
105 if (bkey_extent_is_unwritten(k))
106 return -BCH_ERR_nopromote_unwritten;
108 if (bch2_target_congested(c, opts.promote_target))
109 return -BCH_ERR_nopromote_congested;
111 if (rhashtable_lookup_fast(&c->promote_table, &pos,
113 return -BCH_ERR_nopromote_in_flight;
118 static void promote_free(struct bch_fs *c, struct promote_op *op)
122 bch2_data_update_exit(&op->write);
124 ret = rhashtable_remove_fast(&c->promote_table, &op->hash,
127 bch2_write_ref_put(c, BCH_WRITE_REF_promote);
131 static void promote_done(struct bch_write_op *wop)
133 struct promote_op *op =
134 container_of(wop, struct promote_op, write.op);
135 struct bch_fs *c = op->write.op.c;
137 bch2_time_stats_update(&c->times[BCH_TIME_data_promote],
142 static void promote_start(struct promote_op *op, struct bch_read_bio *rbio)
144 struct bio *bio = &op->write.op.wbio.bio;
146 trace_and_count(op->write.op.c, read_promote, &rbio->bio);
148 /* we now own pages: */
149 BUG_ON(!rbio->bounce);
150 BUG_ON(rbio->bio.bi_vcnt > bio->bi_max_vecs);
152 memcpy(bio->bi_io_vec, rbio->bio.bi_io_vec,
153 sizeof(struct bio_vec) * rbio->bio.bi_vcnt);
154 swap(bio->bi_vcnt, rbio->bio.bi_vcnt);
156 bch2_data_update_read_done(&op->write, rbio->pick.crc);
159 static struct promote_op *__promote_alloc(struct btree_trans *trans,
160 enum btree_id btree_id,
163 struct extent_ptr_decoded *pick,
164 struct bch_io_opts opts,
166 struct bch_read_bio **rbio)
168 struct bch_fs *c = trans->c;
169 struct promote_op *op = NULL;
171 unsigned pages = DIV_ROUND_UP(sectors, PAGE_SECTORS);
174 if (!bch2_write_ref_tryget(c, BCH_WRITE_REF_promote))
175 return ERR_PTR(-BCH_ERR_nopromote_no_writes);
177 op = kzalloc(sizeof(*op) + sizeof(struct bio_vec) * pages, GFP_KERNEL);
179 ret = -BCH_ERR_nopromote_enomem;
183 op->start_time = local_clock();
187 * We don't use the mempool here because extents that aren't
188 * checksummed or compressed can be too big for the mempool:
190 *rbio = kzalloc(sizeof(struct bch_read_bio) +
191 sizeof(struct bio_vec) * pages,
194 ret = -BCH_ERR_nopromote_enomem;
198 rbio_init(&(*rbio)->bio, opts);
199 bio_init(&(*rbio)->bio, NULL, (*rbio)->bio.bi_inline_vecs, pages, 0);
201 if (bch2_bio_alloc_pages(&(*rbio)->bio, sectors << 9, GFP_KERNEL)) {
202 ret = -BCH_ERR_nopromote_enomem;
206 (*rbio)->bounce = true;
207 (*rbio)->split = true;
208 (*rbio)->kmalloc = true;
210 if (rhashtable_lookup_insert_fast(&c->promote_table, &op->hash,
211 bch_promote_params)) {
212 ret = -BCH_ERR_nopromote_in_flight;
216 bio = &op->write.op.wbio.bio;
217 bio_init(bio, NULL, bio->bi_inline_vecs, pages, 0);
219 ret = bch2_data_update_init(trans, NULL, NULL, &op->write,
220 writepoint_hashed((unsigned long) current),
222 (struct data_update_opts) {
223 .target = opts.promote_target,
225 .write_flags = BCH_WRITE_ALLOC_NOWAIT|BCH_WRITE_CACHED,
229 * possible errors: -BCH_ERR_nocow_lock_blocked,
230 * -BCH_ERR_ENOSPC_disk_reservation:
233 BUG_ON(rhashtable_remove_fast(&c->promote_table, &op->hash,
234 bch_promote_params));
238 op->write.op.end_io = promote_done;
243 bio_free_pages(&(*rbio)->bio);
247 bch2_write_ref_put(c, BCH_WRITE_REF_promote);
252 static struct promote_op *promote_alloc(struct btree_trans *trans,
253 struct bvec_iter iter,
255 struct extent_ptr_decoded *pick,
256 struct bch_io_opts opts,
258 struct bch_read_bio **rbio,
262 struct bch_fs *c = trans->c;
263 bool promote_full = *read_full || READ_ONCE(c->promote_whole_extents);
264 /* data might have to be decompressed in the write path: */
265 unsigned sectors = promote_full
266 ? max(pick->crc.compressed_size, pick->crc.live_size)
267 : bvec_iter_sectors(iter);
268 struct bpos pos = promote_full
269 ? bkey_start_pos(k.k)
270 : POS(k.k->p.inode, iter.bi_sector);
271 struct promote_op *promote;
274 ret = should_promote(c, k, pos, opts, flags);
278 promote = __promote_alloc(trans,
279 k.k->type == KEY_TYPE_reflink_v
282 k, pos, pick, opts, sectors, rbio);
283 ret = PTR_ERR_OR_ZERO(promote);
288 *read_full = promote_full;
291 trace_read_nopromote(c, ret);
297 #define READ_RETRY_AVOID 1
303 RBIO_CONTEXT_HIGHPRI,
304 RBIO_CONTEXT_UNBOUND,
307 static inline struct bch_read_bio *
308 bch2_rbio_parent(struct bch_read_bio *rbio)
310 return rbio->split ? rbio->parent : rbio;
314 static void bch2_rbio_punt(struct bch_read_bio *rbio, work_func_t fn,
315 enum rbio_context context,
316 struct workqueue_struct *wq)
318 if (context <= rbio->context) {
321 rbio->work.func = fn;
322 rbio->context = context;
323 queue_work(wq, &rbio->work);
327 static inline struct bch_read_bio *bch2_rbio_free(struct bch_read_bio *rbio)
329 BUG_ON(rbio->bounce && !rbio->split);
332 promote_free(rbio->c, rbio->promote);
333 rbio->promote = NULL;
336 bch2_bio_free_pages_pool(rbio->c, &rbio->bio);
339 struct bch_read_bio *parent = rbio->parent;
353 * Only called on a top level bch_read_bio to complete an entire read request,
356 static void bch2_rbio_done(struct bch_read_bio *rbio)
358 if (rbio->start_time)
359 bch2_time_stats_update(&rbio->c->times[BCH_TIME_data_read],
361 bio_endio(&rbio->bio);
364 static void bch2_read_retry_nodecode(struct bch_fs *c, struct bch_read_bio *rbio,
365 struct bvec_iter bvec_iter,
366 struct bch_io_failures *failed,
369 struct btree_trans *trans = bch2_trans_get(c);
370 struct btree_iter iter;
375 flags &= ~BCH_READ_LAST_FRAGMENT;
376 flags |= BCH_READ_MUST_CLONE;
378 bch2_bkey_buf_init(&sk);
380 bch2_trans_iter_init(trans, &iter, rbio->data_btree,
381 rbio->read_pos, BTREE_ITER_SLOTS);
383 rbio->bio.bi_status = 0;
385 k = bch2_btree_iter_peek_slot(&iter);
389 bch2_bkey_buf_reassemble(&sk, c, k);
390 k = bkey_i_to_s_c(sk.k);
391 bch2_trans_unlock(trans);
393 if (!bch2_bkey_matches_ptr(c, k,
395 rbio->data_pos.offset -
396 rbio->pick.crc.offset)) {
397 /* extent we wanted to read no longer exists: */
402 ret = __bch2_read_extent(trans, rbio, bvec_iter,
405 k, 0, failed, flags);
406 if (ret == READ_RETRY)
411 bch2_rbio_done(rbio);
412 bch2_trans_iter_exit(trans, &iter);
413 bch2_trans_put(trans);
414 bch2_bkey_buf_exit(&sk, c);
417 rbio->bio.bi_status = BLK_STS_IOERR;
421 static void bch2_rbio_retry(struct work_struct *work)
423 struct bch_read_bio *rbio =
424 container_of(work, struct bch_read_bio, work);
425 struct bch_fs *c = rbio->c;
426 struct bvec_iter iter = rbio->bvec_iter;
427 unsigned flags = rbio->flags;
429 .subvol = rbio->subvol,
430 .inum = rbio->read_pos.inode,
432 struct bch_io_failures failed = { .nr = 0 };
434 trace_and_count(c, read_retry, &rbio->bio);
436 if (rbio->retry == READ_RETRY_AVOID)
437 bch2_mark_io_failure(&failed, &rbio->pick);
439 rbio->bio.bi_status = 0;
441 rbio = bch2_rbio_free(rbio);
443 flags |= BCH_READ_IN_RETRY;
444 flags &= ~BCH_READ_MAY_PROMOTE;
446 if (flags & BCH_READ_NODECODE) {
447 bch2_read_retry_nodecode(c, rbio, iter, &failed, flags);
449 flags &= ~BCH_READ_LAST_FRAGMENT;
450 flags |= BCH_READ_MUST_CLONE;
452 __bch2_read(c, rbio, iter, inum, &failed, flags);
456 static void bch2_rbio_error(struct bch_read_bio *rbio, int retry,
461 if (rbio->flags & BCH_READ_IN_RETRY)
464 if (retry == READ_ERR) {
465 rbio = bch2_rbio_free(rbio);
467 rbio->bio.bi_status = error;
468 bch2_rbio_done(rbio);
470 bch2_rbio_punt(rbio, bch2_rbio_retry,
471 RBIO_CONTEXT_UNBOUND, system_unbound_wq);
475 static int __bch2_rbio_narrow_crcs(struct btree_trans *trans,
476 struct bch_read_bio *rbio)
478 struct bch_fs *c = rbio->c;
479 u64 data_offset = rbio->data_pos.offset - rbio->pick.crc.offset;
480 struct bch_extent_crc_unpacked new_crc;
481 struct btree_iter iter;
486 if (crc_is_compressed(rbio->pick.crc))
489 k = bch2_bkey_get_iter(trans, &iter, rbio->data_btree, rbio->data_pos,
490 BTREE_ITER_SLOTS|BTREE_ITER_INTENT);
491 if ((ret = bkey_err(k)))
494 if (bversion_cmp(k.k->version, rbio->version) ||
495 !bch2_bkey_matches_ptr(c, k, rbio->pick.ptr, data_offset))
498 /* Extent was merged? */
499 if (bkey_start_offset(k.k) < data_offset ||
500 k.k->p.offset > data_offset + rbio->pick.crc.uncompressed_size)
503 if (bch2_rechecksum_bio(c, &rbio->bio, rbio->version,
504 rbio->pick.crc, NULL, &new_crc,
505 bkey_start_offset(k.k) - data_offset, k.k->size,
506 rbio->pick.crc.csum_type)) {
507 bch_err(c, "error verifying existing checksum while narrowing checksum (memory corruption?)");
513 * going to be temporarily appending another checksum entry:
515 new = bch2_trans_kmalloc(trans, bkey_bytes(k.k) +
516 sizeof(struct bch_extent_crc128));
517 if ((ret = PTR_ERR_OR_ZERO(new)))
520 bkey_reassemble(new, k);
522 if (!bch2_bkey_narrow_crcs(new, new_crc))
525 ret = bch2_trans_update(trans, &iter, new,
526 BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE);
528 bch2_trans_iter_exit(trans, &iter);
532 static noinline void bch2_rbio_narrow_crcs(struct bch_read_bio *rbio)
534 bch2_trans_do(rbio->c, NULL, NULL, BCH_TRANS_COMMIT_no_enospc,
535 __bch2_rbio_narrow_crcs(trans, rbio));
538 /* Inner part that may run in process context */
539 static void __bch2_read_endio(struct work_struct *work)
541 struct bch_read_bio *rbio =
542 container_of(work, struct bch_read_bio, work);
543 struct bch_fs *c = rbio->c;
544 struct bch_dev *ca = bch_dev_bkey_exists(c, rbio->pick.ptr.dev);
545 struct bio *src = &rbio->bio;
546 struct bio *dst = &bch2_rbio_parent(rbio)->bio;
547 struct bvec_iter dst_iter = rbio->bvec_iter;
548 struct bch_extent_crc_unpacked crc = rbio->pick.crc;
549 struct nonce nonce = extent_nonce(rbio->version, crc);
551 struct bch_csum csum;
554 nofs_flags = memalloc_nofs_save();
556 /* Reset iterator for checksumming and copying bounced data: */
558 src->bi_iter.bi_size = crc.compressed_size << 9;
559 src->bi_iter.bi_idx = 0;
560 src->bi_iter.bi_bvec_done = 0;
562 src->bi_iter = rbio->bvec_iter;
565 csum = bch2_checksum_bio(c, crc.csum_type, nonce, src);
566 if (bch2_crc_cmp(csum, rbio->pick.crc.csum) && !c->opts.no_data_io)
571 * We need to rework the narrow_crcs path to deliver the read completion
572 * first, and then punt to a different workqueue, otherwise we're
573 * holding up reads while doing btree updates which is bad for memory
576 if (unlikely(rbio->narrow_crcs))
577 bch2_rbio_narrow_crcs(rbio);
579 if (rbio->flags & BCH_READ_NODECODE)
582 /* Adjust crc to point to subset of data we want: */
583 crc.offset += rbio->offset_into_extent;
584 crc.live_size = bvec_iter_sectors(rbio->bvec_iter);
586 if (crc_is_compressed(crc)) {
587 ret = bch2_encrypt_bio(c, crc.csum_type, nonce, src);
591 if (bch2_bio_uncompress(c, src, dst, dst_iter, crc) &&
593 goto decompression_err;
595 /* don't need to decrypt the entire bio: */
596 nonce = nonce_add(nonce, crc.offset << 9);
597 bio_advance(src, crc.offset << 9);
599 BUG_ON(src->bi_iter.bi_size < dst_iter.bi_size);
600 src->bi_iter.bi_size = dst_iter.bi_size;
602 ret = bch2_encrypt_bio(c, crc.csum_type, nonce, src);
607 struct bvec_iter src_iter = src->bi_iter;
609 bio_copy_data_iter(dst, &dst_iter, src, &src_iter);
615 * Re encrypt data we decrypted, so it's consistent with
618 ret = bch2_encrypt_bio(c, crc.csum_type, nonce, src);
622 promote_start(rbio->promote, rbio);
623 rbio->promote = NULL;
626 if (likely(!(rbio->flags & BCH_READ_IN_RETRY))) {
627 rbio = bch2_rbio_free(rbio);
628 bch2_rbio_done(rbio);
631 memalloc_nofs_restore(nofs_flags);
635 * Checksum error: if the bio wasn't bounced, we may have been
636 * reading into buffers owned by userspace (that userspace can
637 * scribble over) - retry the read, bouncing it this time:
639 if (!rbio->bounce && (rbio->flags & BCH_READ_USER_MAPPED)) {
640 rbio->flags |= BCH_READ_MUST_BOUNCE;
641 bch2_rbio_error(rbio, READ_RETRY, BLK_STS_IOERR);
645 bch_err_inum_offset_ratelimited(ca,
646 rbio->read_pos.inode,
647 rbio->read_pos.offset << 9,
648 "data checksum error: expected %0llx:%0llx got %0llx:%0llx (type %s)",
649 rbio->pick.crc.csum.hi, rbio->pick.crc.csum.lo,
650 csum.hi, csum.lo, bch2_csum_types[crc.csum_type]);
651 bch2_io_error(ca, BCH_MEMBER_ERROR_checksum);
652 bch2_rbio_error(rbio, READ_RETRY_AVOID, BLK_STS_IOERR);
655 bch_err_inum_offset_ratelimited(c, rbio->read_pos.inode,
656 rbio->read_pos.offset << 9,
657 "decompression error");
658 bch2_rbio_error(rbio, READ_ERR, BLK_STS_IOERR);
661 bch_err_inum_offset_ratelimited(c, rbio->read_pos.inode,
662 rbio->read_pos.offset << 9,
664 bch2_rbio_error(rbio, READ_ERR, BLK_STS_IOERR);
668 static void bch2_read_endio(struct bio *bio)
670 struct bch_read_bio *rbio =
671 container_of(bio, struct bch_read_bio, bio);
672 struct bch_fs *c = rbio->c;
673 struct bch_dev *ca = bch_dev_bkey_exists(c, rbio->pick.ptr.dev);
674 struct workqueue_struct *wq = NULL;
675 enum rbio_context context = RBIO_CONTEXT_NULL;
677 if (rbio->have_ioref) {
678 bch2_latency_acct(ca, rbio->submit_time, READ);
679 percpu_ref_put(&ca->io_ref);
683 rbio->bio.bi_end_io = rbio->end_io;
685 if (bch2_dev_inum_io_err_on(bio->bi_status, ca, BCH_MEMBER_ERROR_read,
686 rbio->read_pos.inode,
687 rbio->read_pos.offset,
688 "data read error: %s",
689 bch2_blk_status_to_str(bio->bi_status))) {
690 bch2_rbio_error(rbio, READ_RETRY_AVOID, bio->bi_status);
694 if (((rbio->flags & BCH_READ_RETRY_IF_STALE) && race_fault()) ||
695 ptr_stale(ca, &rbio->pick.ptr)) {
696 trace_and_count(c, read_reuse_race, &rbio->bio);
698 if (rbio->flags & BCH_READ_RETRY_IF_STALE)
699 bch2_rbio_error(rbio, READ_RETRY, BLK_STS_AGAIN);
701 bch2_rbio_error(rbio, READ_ERR, BLK_STS_AGAIN);
705 if (rbio->narrow_crcs ||
707 crc_is_compressed(rbio->pick.crc) ||
708 bch2_csum_type_is_encryption(rbio->pick.crc.csum_type))
709 context = RBIO_CONTEXT_UNBOUND, wq = system_unbound_wq;
710 else if (rbio->pick.crc.csum_type)
711 context = RBIO_CONTEXT_HIGHPRI, wq = system_highpri_wq;
713 bch2_rbio_punt(rbio, __bch2_read_endio, context, wq);
716 int __bch2_read_indirect_extent(struct btree_trans *trans,
717 unsigned *offset_into_extent,
718 struct bkey_buf *orig_k)
720 struct btree_iter iter;
725 reflink_offset = le64_to_cpu(bkey_i_to_reflink_p(orig_k->k)->v.idx) +
728 k = bch2_bkey_get_iter(trans, &iter, BTREE_ID_reflink,
729 POS(0, reflink_offset), 0);
734 if (k.k->type != KEY_TYPE_reflink_v &&
735 k.k->type != KEY_TYPE_indirect_inline_data) {
736 bch_err_inum_offset_ratelimited(trans->c,
737 orig_k->k->k.p.inode,
738 orig_k->k->k.p.offset << 9,
739 "%llu len %u points to nonexistent indirect extent %llu",
740 orig_k->k->k.p.offset,
743 bch2_inconsistent_error(trans->c);
748 *offset_into_extent = iter.pos.offset - bkey_start_offset(k.k);
749 bch2_bkey_buf_reassemble(orig_k, trans->c, k);
751 bch2_trans_iter_exit(trans, &iter);
755 static noinline void read_from_stale_dirty_pointer(struct btree_trans *trans,
757 struct bch_extent_ptr ptr)
759 struct bch_fs *c = trans->c;
760 struct bch_dev *ca = bch_dev_bkey_exists(c, ptr.dev);
761 struct btree_iter iter;
762 struct printbuf buf = PRINTBUF;
765 bch2_trans_iter_init(trans, &iter, BTREE_ID_alloc,
766 PTR_BUCKET_POS(c, &ptr),
769 prt_printf(&buf, "Attempting to read from stale dirty pointer:");
770 printbuf_indent_add(&buf, 2);
773 bch2_bkey_val_to_text(&buf, c, k);
776 prt_printf(&buf, "memory gen: %u", *bucket_gen(ca, iter.pos.offset));
778 ret = lockrestart_do(trans, bkey_err(k = bch2_btree_iter_peek_slot(&iter)));
781 bch2_bkey_val_to_text(&buf, c, k);
784 bch2_fs_inconsistent(c, "%s", buf.buf);
786 bch2_trans_iter_exit(trans, &iter);
790 int __bch2_read_extent(struct btree_trans *trans, struct bch_read_bio *orig,
791 struct bvec_iter iter, struct bpos read_pos,
792 enum btree_id data_btree, struct bkey_s_c k,
793 unsigned offset_into_extent,
794 struct bch_io_failures *failed, unsigned flags)
796 struct bch_fs *c = trans->c;
797 struct extent_ptr_decoded pick;
798 struct bch_read_bio *rbio = NULL;
799 struct bch_dev *ca = NULL;
800 struct promote_op *promote = NULL;
801 bool bounce = false, read_full = false, narrow_crcs = false;
802 struct bpos data_pos = bkey_start_pos(k.k);
805 if (bkey_extent_is_inline_data(k.k)) {
806 unsigned bytes = min_t(unsigned, iter.bi_size,
807 bkey_inline_data_bytes(k.k));
809 swap(iter.bi_size, bytes);
810 memcpy_to_bio(&orig->bio, iter, bkey_inline_data_p(k));
811 swap(iter.bi_size, bytes);
812 bio_advance_iter(&orig->bio, &iter, bytes);
813 zero_fill_bio_iter(&orig->bio, iter);
817 pick_ret = bch2_bkey_pick_read_device(c, k, failed, &pick);
819 /* hole or reservation - just zero fill: */
824 bch_err_inum_offset_ratelimited(c,
825 read_pos.inode, read_pos.offset << 9,
826 "no device to read from");
830 ca = bch_dev_bkey_exists(c, pick.ptr.dev);
833 * Stale dirty pointers are treated as IO errors, but @failed isn't
834 * allocated unless we're in the retry path - so if we're not in the
835 * retry path, don't check here, it'll be caught in bch2_read_endio()
836 * and we'll end up in the retry path:
838 if ((flags & BCH_READ_IN_RETRY) &&
840 unlikely(ptr_stale(ca, &pick.ptr))) {
841 read_from_stale_dirty_pointer(trans, k, pick.ptr);
842 bch2_mark_io_failure(failed, &pick);
847 * Unlock the iterator while the btree node's lock is still in
848 * cache, before doing the IO:
850 bch2_trans_unlock(trans);
852 if (flags & BCH_READ_NODECODE) {
854 * can happen if we retry, and the extent we were going to read
855 * has been merged in the meantime:
857 if (pick.crc.compressed_size > orig->bio.bi_vcnt * PAGE_SECTORS)
860 iter.bi_size = pick.crc.compressed_size << 9;
864 if (!(flags & BCH_READ_LAST_FRAGMENT) ||
865 bio_flagged(&orig->bio, BIO_CHAIN))
866 flags |= BCH_READ_MUST_CLONE;
868 narrow_crcs = !(flags & BCH_READ_IN_RETRY) &&
869 bch2_can_narrow_extent_crcs(k, pick.crc);
871 if (narrow_crcs && (flags & BCH_READ_USER_MAPPED))
872 flags |= BCH_READ_MUST_BOUNCE;
874 EBUG_ON(offset_into_extent + bvec_iter_sectors(iter) > k.k->size);
876 if (crc_is_compressed(pick.crc) ||
877 (pick.crc.csum_type != BCH_CSUM_none &&
878 (bvec_iter_sectors(iter) != pick.crc.uncompressed_size ||
879 (bch2_csum_type_is_encryption(pick.crc.csum_type) &&
880 (flags & BCH_READ_USER_MAPPED)) ||
881 (flags & BCH_READ_MUST_BOUNCE)))) {
886 if (orig->opts.promote_target)
887 promote = promote_alloc(trans, iter, k, &pick, orig->opts, flags,
888 &rbio, &bounce, &read_full);
891 EBUG_ON(crc_is_compressed(pick.crc));
892 EBUG_ON(pick.crc.csum_type &&
893 (bvec_iter_sectors(iter) != pick.crc.uncompressed_size ||
894 bvec_iter_sectors(iter) != pick.crc.live_size ||
896 offset_into_extent));
898 data_pos.offset += offset_into_extent;
899 pick.ptr.offset += pick.crc.offset +
901 offset_into_extent = 0;
902 pick.crc.compressed_size = bvec_iter_sectors(iter);
903 pick.crc.uncompressed_size = bvec_iter_sectors(iter);
905 pick.crc.live_size = bvec_iter_sectors(iter);
910 * promote already allocated bounce rbio:
911 * promote needs to allocate a bio big enough for uncompressing
912 * data in the write path, but we're not going to use it all
915 EBUG_ON(rbio->bio.bi_iter.bi_size <
916 pick.crc.compressed_size << 9);
917 rbio->bio.bi_iter.bi_size =
918 pick.crc.compressed_size << 9;
920 unsigned sectors = pick.crc.compressed_size;
922 rbio = rbio_init(bio_alloc_bioset(NULL,
923 DIV_ROUND_UP(sectors, PAGE_SECTORS),
929 bch2_bio_alloc_pages_pool(c, &rbio->bio, sectors << 9);
932 } else if (flags & BCH_READ_MUST_CLONE) {
934 * Have to clone if there were any splits, due to error
935 * reporting issues (if a split errored, and retrying didn't
936 * work, when it reports the error to its parent (us) we don't
937 * know if the error was from our bio, and we should retry, or
938 * from the whole bio, in which case we don't want to retry and
941 rbio = rbio_init(bio_alloc_clone(NULL, &orig->bio, GFP_NOFS,
944 rbio->bio.bi_iter = iter;
948 rbio->bio.bi_iter = iter;
949 EBUG_ON(bio_flagged(&rbio->bio, BIO_CHAIN));
952 EBUG_ON(bio_sectors(&rbio->bio) != pick.crc.compressed_size);
955 rbio->submit_time = local_clock();
959 rbio->end_io = orig->bio.bi_end_io;
960 rbio->bvec_iter = iter;
961 rbio->offset_into_extent= offset_into_extent;
963 rbio->have_ioref = pick_ret > 0 && bch2_dev_get_ioref(ca, READ);
964 rbio->narrow_crcs = narrow_crcs;
968 /* XXX: only initialize this if needed */
969 rbio->devs_have = bch2_bkey_devs(k);
971 rbio->subvol = orig->subvol;
972 rbio->read_pos = read_pos;
973 rbio->data_btree = data_btree;
974 rbio->data_pos = data_pos;
975 rbio->version = k.k->version;
976 rbio->promote = promote;
977 INIT_WORK(&rbio->work, NULL);
979 rbio->bio.bi_opf = orig->bio.bi_opf;
980 rbio->bio.bi_iter.bi_sector = pick.ptr.offset;
981 rbio->bio.bi_end_io = bch2_read_endio;
984 trace_and_count(c, read_bounce, &rbio->bio);
986 this_cpu_add(c->counters[BCH_COUNTER_io_read], bio_sectors(&rbio->bio));
987 bch2_increment_clock(c, bio_sectors(&rbio->bio), READ);
990 * If it's being moved internally, we don't want to flag it as a cache
993 if (pick.ptr.cached && !(flags & BCH_READ_NODECODE))
994 bch2_bucket_io_time_reset(trans, pick.ptr.dev,
995 PTR_BUCKET_NR(ca, &pick.ptr), READ);
997 if (!(flags & (BCH_READ_IN_RETRY|BCH_READ_LAST_FRAGMENT))) {
998 bio_inc_remaining(&orig->bio);
999 trace_and_count(c, read_split, &orig->bio);
1002 if (!rbio->pick.idx) {
1003 if (!rbio->have_ioref) {
1004 bch_err_inum_offset_ratelimited(c,
1006 read_pos.offset << 9,
1007 "no device to read from");
1008 bch2_rbio_error(rbio, READ_RETRY_AVOID, BLK_STS_IOERR);
1012 this_cpu_add(ca->io_done->sectors[READ][BCH_DATA_user],
1013 bio_sectors(&rbio->bio));
1014 bio_set_dev(&rbio->bio, ca->disk_sb.bdev);
1016 if (unlikely(c->opts.no_data_io)) {
1017 if (likely(!(flags & BCH_READ_IN_RETRY)))
1018 bio_endio(&rbio->bio);
1020 if (likely(!(flags & BCH_READ_IN_RETRY)))
1021 submit_bio(&rbio->bio);
1023 submit_bio_wait(&rbio->bio);
1027 * We just submitted IO which may block, we expect relock fail
1028 * events and shouldn't count them:
1030 trans->notrace_relock_fail = true;
1032 /* Attempting reconstruct read: */
1033 if (bch2_ec_read_extent(trans, rbio)) {
1034 bch2_rbio_error(rbio, READ_RETRY_AVOID, BLK_STS_IOERR);
1038 if (likely(!(flags & BCH_READ_IN_RETRY)))
1039 bio_endio(&rbio->bio);
1042 if (likely(!(flags & BCH_READ_IN_RETRY))) {
1047 rbio->context = RBIO_CONTEXT_UNBOUND;
1048 bch2_read_endio(&rbio->bio);
1051 rbio = bch2_rbio_free(rbio);
1053 if (ret == READ_RETRY_AVOID) {
1054 bch2_mark_io_failure(failed, &pick);
1065 if (flags & BCH_READ_IN_RETRY)
1068 orig->bio.bi_status = BLK_STS_IOERR;
1073 * won't normally happen in the BCH_READ_NODECODE
1074 * (bch2_move_extent()) path, but if we retry and the extent we wanted
1075 * to read no longer exists we have to signal that:
1077 if (flags & BCH_READ_NODECODE)
1080 zero_fill_bio_iter(&orig->bio, iter);
1082 if (flags & BCH_READ_LAST_FRAGMENT)
1083 bch2_rbio_done(orig);
1087 void __bch2_read(struct bch_fs *c, struct bch_read_bio *rbio,
1088 struct bvec_iter bvec_iter, subvol_inum inum,
1089 struct bch_io_failures *failed, unsigned flags)
1091 struct btree_trans *trans = bch2_trans_get(c);
1092 struct btree_iter iter;
1098 BUG_ON(flags & BCH_READ_NODECODE);
1100 bch2_bkey_buf_init(&sk);
1102 bch2_trans_begin(trans);
1103 iter = (struct btree_iter) { NULL };
1105 ret = bch2_subvolume_get_snapshot(trans, inum.subvol, &snapshot);
1109 bch2_trans_iter_init(trans, &iter, BTREE_ID_extents,
1110 SPOS(inum.inum, bvec_iter.bi_sector, snapshot),
1113 unsigned bytes, sectors, offset_into_extent;
1114 enum btree_id data_btree = BTREE_ID_extents;
1117 * read_extent -> io_time_reset may cause a transaction restart
1118 * without returning an error, we need to check for that here:
1120 ret = bch2_trans_relock(trans);
1124 bch2_btree_iter_set_pos(&iter,
1125 POS(inum.inum, bvec_iter.bi_sector));
1127 k = bch2_btree_iter_peek_slot(&iter);
1132 offset_into_extent = iter.pos.offset -
1133 bkey_start_offset(k.k);
1134 sectors = k.k->size - offset_into_extent;
1136 bch2_bkey_buf_reassemble(&sk, c, k);
1138 ret = bch2_read_indirect_extent(trans, &data_btree,
1139 &offset_into_extent, &sk);
1143 k = bkey_i_to_s_c(sk.k);
1146 * With indirect extents, the amount of data to read is the min
1147 * of the original extent and the indirect extent:
1149 sectors = min(sectors, k.k->size - offset_into_extent);
1151 bytes = min(sectors, bvec_iter_sectors(bvec_iter)) << 9;
1152 swap(bvec_iter.bi_size, bytes);
1154 if (bvec_iter.bi_size == bytes)
1155 flags |= BCH_READ_LAST_FRAGMENT;
1157 ret = __bch2_read_extent(trans, rbio, bvec_iter, iter.pos,
1159 offset_into_extent, failed, flags);
1163 if (flags & BCH_READ_LAST_FRAGMENT)
1166 swap(bvec_iter.bi_size, bytes);
1167 bio_advance_iter(&rbio->bio, &bvec_iter, bytes);
1169 ret = btree_trans_too_many_iters(trans);
1174 bch2_trans_iter_exit(trans, &iter);
1176 if (bch2_err_matches(ret, BCH_ERR_transaction_restart) ||
1177 ret == READ_RETRY ||
1178 ret == READ_RETRY_AVOID)
1181 bch2_trans_put(trans);
1182 bch2_bkey_buf_exit(&sk, c);
1185 bch_err_inum_offset_ratelimited(c, inum.inum,
1186 bvec_iter.bi_sector << 9,
1187 "read error %i from btree lookup", ret);
1188 rbio->bio.bi_status = BLK_STS_IOERR;
1189 bch2_rbio_done(rbio);
1193 void bch2_fs_io_read_exit(struct bch_fs *c)
1195 if (c->promote_table.tbl)
1196 rhashtable_destroy(&c->promote_table);
1197 bioset_exit(&c->bio_read_split);
1198 bioset_exit(&c->bio_read);
1201 int bch2_fs_io_read_init(struct bch_fs *c)
1203 if (bioset_init(&c->bio_read, 1, offsetof(struct bch_read_bio, bio),
1205 return -BCH_ERR_ENOMEM_bio_read_init;
1207 if (bioset_init(&c->bio_read_split, 1, offsetof(struct bch_read_bio, bio),
1209 return -BCH_ERR_ENOMEM_bio_read_split_init;
1211 if (rhashtable_init(&c->promote_table, &bch_promote_params))
1212 return -BCH_ERR_ENOMEM_promote_table_init;