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"
14 #include "btree_update.h"
20 #include "disk_groups.h"
23 #include "extent_update.h"
29 #include "rebalance.h"
30 #include "subvolume.h"
34 #include <linux/blkdev.h>
35 #include <linux/random.h>
36 #include <linux/sched/mm.h>
38 #include <trace/events/bcachefs.h>
40 const char *bch2_blk_status_to_str(blk_status_t status)
42 if (status == BLK_STS_REMOVED)
43 return "device removed";
44 return blk_status_to_str(status);
47 static bool bch2_target_congested(struct bch_fs *c, u16 target)
49 const struct bch_devs_mask *devs;
50 unsigned d, nr = 0, total = 0;
51 u64 now = local_clock(), last;
59 devs = bch2_target_to_mask(c, target) ?:
60 &c->rw_devs[BCH_DATA_user];
62 for_each_set_bit(d, devs->d, BCH_SB_MEMBERS_MAX) {
63 ca = rcu_dereference(c->devs[d]);
67 congested = atomic_read(&ca->congested);
68 last = READ_ONCE(ca->congested_last);
69 if (time_after64(now, last))
70 congested -= (now - last) >> 12;
72 total += max(congested, 0LL);
77 return bch2_rand_range(nr * CONGESTED_MAX) < total;
80 static inline void bch2_congested_acct(struct bch_dev *ca, u64 io_latency,
84 ca->io_latency[rw].quantiles.entries[QUANTILE_IDX(1)].m;
85 /* ideally we'd be taking into account the device's variance here: */
86 u64 latency_threshold = latency_capable << (rw == READ ? 2 : 3);
87 s64 latency_over = io_latency - latency_threshold;
89 if (latency_threshold && latency_over > 0) {
91 * bump up congested by approximately latency_over * 4 /
92 * latency_threshold - we don't need much accuracy here so don't
93 * bother with the divide:
95 if (atomic_read(&ca->congested) < CONGESTED_MAX)
96 atomic_add(latency_over >>
97 max_t(int, ilog2(latency_threshold) - 2, 0),
100 ca->congested_last = now;
101 } else if (atomic_read(&ca->congested) > 0) {
102 atomic_dec(&ca->congested);
106 void bch2_latency_acct(struct bch_dev *ca, u64 submit_time, int rw)
108 atomic64_t *latency = &ca->cur_latency[rw];
109 u64 now = local_clock();
110 u64 io_latency = time_after64(now, submit_time)
113 u64 old, new, v = atomic64_read(latency);
119 * If the io latency was reasonably close to the current
120 * latency, skip doing the update and atomic operation - most of
123 if (abs((int) (old - io_latency)) < (old >> 1) &&
127 new = ewma_add(old, io_latency, 5);
128 } while ((v = atomic64_cmpxchg(latency, old, new)) != old);
130 bch2_congested_acct(ca, io_latency, now, rw);
132 __bch2_time_stats_update(&ca->io_latency[rw], submit_time, now);
135 /* Allocate, free from mempool: */
137 void bch2_bio_free_pages_pool(struct bch_fs *c, struct bio *bio)
139 struct bvec_iter_all iter;
142 bio_for_each_segment_all(bv, bio, iter)
143 if (bv->bv_page != ZERO_PAGE(0))
144 mempool_free(bv->bv_page, &c->bio_bounce_pages);
148 static struct page *__bio_alloc_page_pool(struct bch_fs *c, bool *using_mempool)
152 if (likely(!*using_mempool)) {
153 page = alloc_page(GFP_NOIO);
154 if (unlikely(!page)) {
155 mutex_lock(&c->bio_bounce_pages_lock);
156 *using_mempool = true;
162 page = mempool_alloc(&c->bio_bounce_pages, GFP_NOIO);
168 void bch2_bio_alloc_pages_pool(struct bch_fs *c, struct bio *bio,
171 bool using_mempool = false;
174 struct page *page = __bio_alloc_page_pool(c, &using_mempool);
175 unsigned len = min_t(size_t, PAGE_SIZE, size);
177 BUG_ON(!bio_add_page(bio, page, len, 0));
182 mutex_unlock(&c->bio_bounce_pages_lock);
185 /* Extent update path: */
187 int bch2_sum_sector_overwrites(struct btree_trans *trans,
188 struct btree_iter *extent_iter,
190 bool *usage_increasing,
191 s64 *i_sectors_delta,
192 s64 *disk_sectors_delta)
194 struct bch_fs *c = trans->c;
195 struct btree_iter iter;
197 unsigned new_replicas = bch2_bkey_replicas(c, bkey_i_to_s_c(new));
198 bool new_compressed = bch2_bkey_sectors_compressed(bkey_i_to_s_c(new));
201 *usage_increasing = false;
202 *i_sectors_delta = 0;
203 *disk_sectors_delta = 0;
205 bch2_trans_copy_iter(&iter, extent_iter);
207 for_each_btree_key_continue_norestart(iter, BTREE_ITER_SLOTS, old, ret) {
208 s64 sectors = min(new->k.p.offset, old.k->p.offset) -
209 max(bkey_start_offset(&new->k),
210 bkey_start_offset(old.k));
212 *i_sectors_delta += sectors *
213 (bkey_extent_is_allocation(&new->k) -
214 bkey_extent_is_allocation(old.k));
216 *disk_sectors_delta += sectors * bch2_bkey_nr_ptrs_allocated(bkey_i_to_s_c(new));
217 *disk_sectors_delta -= new->k.p.snapshot == old.k->p.snapshot
218 ? sectors * bch2_bkey_nr_ptrs_fully_allocated(old)
221 if (!*usage_increasing &&
222 (new->k.p.snapshot != old.k->p.snapshot ||
223 new_replicas > bch2_bkey_replicas(c, old) ||
224 (!new_compressed && bch2_bkey_sectors_compressed(old))))
225 *usage_increasing = true;
227 if (bkey_cmp(old.k->p, new->k.p) >= 0)
231 bch2_trans_iter_exit(trans, &iter);
235 int bch2_extent_update(struct btree_trans *trans,
237 struct btree_iter *iter,
239 struct disk_reservation *disk_res,
242 s64 *i_sectors_delta_total,
245 struct btree_iter inode_iter;
246 struct bch_inode_unpacked inode_u;
247 struct bpos next_pos;
248 bool usage_increasing;
249 s64 i_sectors_delta = 0, disk_sectors_delta = 0;
253 * This traverses us the iterator without changing iter->path->pos to
254 * search_key() (which is pos + 1 for extents): we want there to be a
255 * path already traversed at iter->pos because
256 * bch2_trans_extent_update() will use it to attempt extent merging
258 ret = __bch2_btree_iter_traverse(iter);
262 ret = bch2_extent_trim_atomic(trans, iter, k);
266 new_i_size = min(k->k.p.offset << 9, new_i_size);
269 ret = bch2_sum_sector_overwrites(trans, iter, k,
272 &disk_sectors_delta);
277 disk_sectors_delta > (s64) disk_res->sectors) {
278 ret = bch2_disk_reservation_add(trans->c, disk_res,
279 disk_sectors_delta - disk_res->sectors,
280 !check_enospc || !usage_increasing
281 ? BCH_DISK_RESERVATION_NOFAIL : 0);
286 ret = bch2_inode_peek(trans, &inode_iter, &inode_u, inum,
291 if (!(inode_u.bi_flags & BCH_INODE_I_SIZE_DIRTY) &&
292 new_i_size > inode_u.bi_size)
293 inode_u.bi_size = new_i_size;
295 inode_u.bi_sectors += i_sectors_delta;
297 ret = bch2_trans_update(trans, iter, k, 0) ?:
298 bch2_inode_write(trans, &inode_iter, &inode_u) ?:
299 bch2_trans_commit(trans, disk_res, journal_seq,
300 BTREE_INSERT_NOCHECK_RW|
301 BTREE_INSERT_NOFAIL);
302 bch2_trans_iter_exit(trans, &inode_iter);
307 if (i_sectors_delta_total)
308 *i_sectors_delta_total += i_sectors_delta;
309 bch2_btree_iter_set_pos(iter, next_pos);
315 * Returns -BCH_ERR_transacton_restart if we had to drop locks:
317 int bch2_fpunch_at(struct btree_trans *trans, struct btree_iter *iter,
318 subvol_inum inum, u64 end,
319 s64 *i_sectors_delta)
321 struct bch_fs *c = trans->c;
322 unsigned max_sectors = KEY_SIZE_MAX & (~0 << c->block_bits);
323 struct bpos end_pos = POS(inum.inum, end);
325 int ret = 0, ret2 = 0;
329 bch2_err_matches(ret, BCH_ERR_transaction_restart)) {
330 struct disk_reservation disk_res =
331 bch2_disk_reservation_init(c, 0);
332 struct bkey_i delete;
337 bch2_trans_begin(trans);
339 ret = bch2_subvolume_get_snapshot(trans, inum.subvol, &snapshot);
343 bch2_btree_iter_set_snapshot(iter, snapshot);
345 k = bch2_btree_iter_peek(iter);
346 if (bkey_cmp(iter->pos, end_pos) >= 0) {
347 bch2_btree_iter_set_pos(iter, end_pos);
355 bkey_init(&delete.k);
356 delete.k.p = iter->pos;
358 /* create the biggest key we can */
359 bch2_key_resize(&delete.k, max_sectors);
360 bch2_cut_back(end_pos, &delete);
362 ret = bch2_extent_update(trans, inum, iter, &delete,
364 0, i_sectors_delta, false);
365 bch2_disk_reservation_put(c, &disk_res);
371 int bch2_fpunch(struct bch_fs *c, subvol_inum inum, u64 start, u64 end,
372 s64 *i_sectors_delta)
374 struct btree_trans trans;
375 struct btree_iter iter;
378 bch2_trans_init(&trans, c, BTREE_ITER_MAX, 1024);
379 bch2_trans_iter_init(&trans, &iter, BTREE_ID_extents,
380 POS(inum.inum, start),
383 ret = bch2_fpunch_at(&trans, &iter, inum, end, i_sectors_delta);
385 bch2_trans_iter_exit(&trans, &iter);
386 bch2_trans_exit(&trans);
388 if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
394 int bch2_write_index_default(struct bch_write_op *op)
396 struct bch_fs *c = op->c;
398 struct open_bucket *ec_ob = ec_open_bucket(c, &op->open_buckets);
399 struct keylist *keys = &op->insert_keys;
400 struct bkey_i *k = bch2_keylist_front(keys);
401 struct btree_trans trans;
402 struct btree_iter iter;
404 .subvol = op->subvol,
405 .inum = k->k.p.inode,
409 BUG_ON(!inum.subvol);
411 bch2_bkey_buf_init(&sk);
412 bch2_trans_init(&trans, c, BTREE_ITER_MAX, 1024);
415 bch2_trans_begin(&trans);
417 k = bch2_keylist_front(keys);
418 bch2_bkey_buf_copy(&sk, c, k);
420 ret = bch2_subvolume_get_snapshot(&trans, inum.subvol,
421 &sk.k->k.p.snapshot);
422 if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
427 bch2_trans_iter_init(&trans, &iter, BTREE_ID_extents,
428 bkey_start_pos(&sk.k->k),
429 BTREE_ITER_SLOTS|BTREE_ITER_INTENT);
431 ret = bch2_extent_update(&trans, inum, &iter, sk.k,
432 &op->res, op_journal_seq(op),
433 op->new_i_size, &op->i_sectors_delta,
434 op->flags & BCH_WRITE_CHECK_ENOSPC);
435 bch2_trans_iter_exit(&trans, &iter);
437 if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
443 bch2_ob_add_backpointer(c, ec_ob, &sk.k->k);
445 if (bkey_cmp(iter.pos, k->k.p) >= 0)
446 bch2_keylist_pop_front(&op->insert_keys);
448 bch2_cut_front(iter.pos, k);
449 } while (!bch2_keylist_empty(keys));
451 bch2_trans_exit(&trans);
452 bch2_bkey_buf_exit(&sk, c);
459 void bch2_submit_wbio_replicas(struct bch_write_bio *wbio, struct bch_fs *c,
460 enum bch_data_type type,
461 const struct bkey_i *k)
463 struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(bkey_i_to_s_c(k));
464 const struct bch_extent_ptr *ptr;
465 struct bch_write_bio *n;
468 BUG_ON(c->opts.nochanges);
470 bkey_for_each_ptr(ptrs, ptr) {
471 BUG_ON(ptr->dev >= BCH_SB_MEMBERS_MAX ||
474 ca = bch_dev_bkey_exists(c, ptr->dev);
476 if (to_entry(ptr + 1) < ptrs.end) {
477 n = to_wbio(bio_alloc_clone(NULL, &wbio->bio,
478 GFP_NOIO, &ca->replica_set));
480 n->bio.bi_end_io = wbio->bio.bi_end_io;
481 n->bio.bi_private = wbio->bio.bi_private;
486 n->bio.bi_opf = wbio->bio.bi_opf;
487 bio_inc_remaining(&wbio->bio);
495 n->have_ioref = bch2_dev_get_ioref(ca,
496 type == BCH_DATA_btree ? READ : WRITE);
497 n->submit_time = local_clock();
498 n->bio.bi_iter.bi_sector = ptr->offset;
500 if (likely(n->have_ioref)) {
501 this_cpu_add(ca->io_done->sectors[WRITE][type],
502 bio_sectors(&n->bio));
504 bio_set_dev(&n->bio, ca->disk_sb.bdev);
507 n->bio.bi_status = BLK_STS_REMOVED;
513 static void __bch2_write(struct closure *);
515 static void bch2_write_done(struct closure *cl)
517 struct bch_write_op *op = container_of(cl, struct bch_write_op, cl);
518 struct bch_fs *c = op->c;
520 if (!op->error && (op->flags & BCH_WRITE_FLUSH))
521 op->error = bch2_journal_error(&c->journal);
523 bch2_disk_reservation_put(c, &op->res);
524 percpu_ref_put(&c->writes);
525 bch2_keylist_free(&op->insert_keys, op->inline_keys);
527 bch2_time_stats_update(&c->times[BCH_TIME_data_write], op->start_time);
531 closure_debug_destroy(cl);
539 * bch_write_index - after a write, update index to point to new data
541 static void __bch2_write_index(struct bch_write_op *op)
543 struct bch_fs *c = op->c;
544 struct keylist *keys = &op->insert_keys;
545 struct bch_extent_ptr *ptr;
546 struct bkey_i *src, *dst = keys->keys, *n, *k;
550 for (src = keys->keys; src != keys->top; src = n) {
553 if (bkey_extent_is_direct_data(&src->k)) {
554 bch2_bkey_drop_ptrs(bkey_i_to_s(src), ptr,
555 test_bit(ptr->dev, op->failed.d));
557 if (!bch2_bkey_nr_ptrs(bkey_i_to_s_c(src))) {
564 memmove_u64s_down(dst, src, src->u64s);
565 dst = bkey_next(dst);
571 * probably not the ideal place to hook this in, but I don't
572 * particularly want to plumb io_opts all the way through the btree
573 * update stack right now
575 for_each_keylist_key(keys, k) {
576 bch2_rebalance_add_key(c, bkey_i_to_s_c(k), &op->opts);
578 if (bch2_bkey_is_incompressible(bkey_i_to_s_c(k)))
579 bch2_check_set_feature(op->c, BCH_FEATURE_incompressible);
583 if (!bch2_keylist_empty(keys)) {
584 u64 sectors_start = keylist_sectors(keys);
585 int ret = op->index_update_fn(op);
587 BUG_ON(bch2_err_matches(ret, BCH_ERR_transaction_restart));
588 BUG_ON(keylist_sectors(keys) && !ret);
590 op->written += sectors_start - keylist_sectors(keys);
593 bch_err_inum_ratelimited(c, op->pos.inode,
594 "write error while doing btree update: %s", bch2_err_str(ret));
599 /* If some a bucket wasn't written, we can't erasure code it: */
600 for_each_set_bit(dev, op->failed.d, BCH_SB_MEMBERS_MAX)
601 bch2_open_bucket_write_error(c, &op->open_buckets, dev);
603 bch2_open_buckets_put(c, &op->open_buckets);
606 keys->top = keys->keys;
611 static void bch2_write_index(struct closure *cl)
613 struct bch_write_op *op = container_of(cl, struct bch_write_op, cl);
614 struct bch_fs *c = op->c;
616 __bch2_write_index(op);
618 if (!(op->flags & BCH_WRITE_DONE)) {
619 continue_at(cl, __bch2_write, index_update_wq(op));
620 } else if (!op->error && (op->flags & BCH_WRITE_FLUSH)) {
621 bch2_journal_flush_seq_async(&c->journal,
624 continue_at(cl, bch2_write_done, index_update_wq(op));
626 continue_at_nobarrier(cl, bch2_write_done, NULL);
630 static void bch2_write_endio(struct bio *bio)
632 struct closure *cl = bio->bi_private;
633 struct bch_write_op *op = container_of(cl, struct bch_write_op, cl);
634 struct bch_write_bio *wbio = to_wbio(bio);
635 struct bch_write_bio *parent = wbio->split ? wbio->parent : NULL;
636 struct bch_fs *c = wbio->c;
637 struct bch_dev *ca = bch_dev_bkey_exists(c, wbio->dev);
639 if (bch2_dev_inum_io_err_on(bio->bi_status, ca,
641 op->pos.offset - bio_sectors(bio), /* XXX definitely wrong */
642 "data write error: %s",
643 bch2_blk_status_to_str(bio->bi_status)))
644 set_bit(wbio->dev, op->failed.d);
646 if (wbio->have_ioref) {
647 bch2_latency_acct(ca, wbio->submit_time, WRITE);
648 percpu_ref_put(&ca->io_ref);
652 bch2_bio_free_pages_pool(c, bio);
658 bio_endio(&parent->bio);
659 else if (!(op->flags & BCH_WRITE_SKIP_CLOSURE_PUT))
662 continue_at_nobarrier(cl, bch2_write_index, index_update_wq(op));
665 static void init_append_extent(struct bch_write_op *op,
666 struct write_point *wp,
667 struct bversion version,
668 struct bch_extent_crc_unpacked crc)
670 struct bch_fs *c = op->c;
671 struct bkey_i_extent *e;
673 op->pos.offset += crc.uncompressed_size;
675 e = bkey_extent_init(op->insert_keys.top);
677 e->k.size = crc.uncompressed_size;
678 e->k.version = version;
681 crc.compression_type ||
683 bch2_extent_crc_append(&e->k_i, crc);
685 bch2_alloc_sectors_append_ptrs(c, wp, &e->k_i, crc.compressed_size,
686 op->flags & BCH_WRITE_CACHED);
688 bch2_keylist_push(&op->insert_keys);
691 static struct bio *bch2_write_bio_alloc(struct bch_fs *c,
692 struct write_point *wp,
694 bool *page_alloc_failed,
697 struct bch_write_bio *wbio;
699 unsigned output_available =
700 min(wp->sectors_free << 9, src->bi_iter.bi_size);
701 unsigned pages = DIV_ROUND_UP(output_available +
703 ? ((unsigned long) buf & (PAGE_SIZE - 1))
706 pages = min(pages, BIO_MAX_VECS);
708 bio = bio_alloc_bioset(NULL, pages, 0,
709 GFP_NOIO, &c->bio_write);
710 wbio = wbio_init(bio);
711 wbio->put_bio = true;
712 /* copy WRITE_SYNC flag */
713 wbio->bio.bi_opf = src->bi_opf;
716 bch2_bio_map(bio, buf, output_available);
723 * We can't use mempool for more than c->sb.encoded_extent_max
724 * worth of pages, but we'd like to allocate more if we can:
726 bch2_bio_alloc_pages_pool(c, bio,
727 min_t(unsigned, output_available,
728 c->opts.encoded_extent_max));
730 if (bio->bi_iter.bi_size < output_available)
732 bch2_bio_alloc_pages(bio,
734 bio->bi_iter.bi_size,
740 static int bch2_write_rechecksum(struct bch_fs *c,
741 struct bch_write_op *op,
742 unsigned new_csum_type)
744 struct bio *bio = &op->wbio.bio;
745 struct bch_extent_crc_unpacked new_crc;
748 /* bch2_rechecksum_bio() can't encrypt or decrypt data: */
750 if (bch2_csum_type_is_encryption(op->crc.csum_type) !=
751 bch2_csum_type_is_encryption(new_csum_type))
752 new_csum_type = op->crc.csum_type;
754 ret = bch2_rechecksum_bio(c, bio, op->version, op->crc,
756 op->crc.offset, op->crc.live_size,
761 bio_advance(bio, op->crc.offset << 9);
762 bio->bi_iter.bi_size = op->crc.live_size << 9;
767 static int bch2_write_decrypt(struct bch_write_op *op)
769 struct bch_fs *c = op->c;
770 struct nonce nonce = extent_nonce(op->version, op->crc);
771 struct bch_csum csum;
774 if (!bch2_csum_type_is_encryption(op->crc.csum_type))
778 * If we need to decrypt data in the write path, we'll no longer be able
779 * to verify the existing checksum (poly1305 mac, in this case) after
780 * it's decrypted - this is the last point we'll be able to reverify the
783 csum = bch2_checksum_bio(c, op->crc.csum_type, nonce, &op->wbio.bio);
784 if (bch2_crc_cmp(op->crc.csum, csum))
787 ret = bch2_encrypt_bio(c, op->crc.csum_type, nonce, &op->wbio.bio);
788 op->crc.csum_type = 0;
789 op->crc.csum = (struct bch_csum) { 0, 0 };
793 static enum prep_encoded_ret {
796 PREP_ENCODED_CHECKSUM_ERR,
797 PREP_ENCODED_DO_WRITE,
798 } bch2_write_prep_encoded_data(struct bch_write_op *op, struct write_point *wp)
800 struct bch_fs *c = op->c;
801 struct bio *bio = &op->wbio.bio;
803 if (!(op->flags & BCH_WRITE_DATA_ENCODED))
804 return PREP_ENCODED_OK;
806 BUG_ON(bio_sectors(bio) != op->crc.compressed_size);
808 /* Can we just write the entire extent as is? */
809 if (op->crc.uncompressed_size == op->crc.live_size &&
810 op->crc.compressed_size <= wp->sectors_free &&
811 (op->crc.compression_type == op->compression_type ||
812 op->incompressible)) {
813 if (!crc_is_compressed(op->crc) &&
814 op->csum_type != op->crc.csum_type &&
815 bch2_write_rechecksum(c, op, op->csum_type))
816 return PREP_ENCODED_CHECKSUM_ERR;
818 return PREP_ENCODED_DO_WRITE;
822 * If the data is compressed and we couldn't write the entire extent as
823 * is, we have to decompress it:
825 if (crc_is_compressed(op->crc)) {
826 struct bch_csum csum;
828 if (bch2_write_decrypt(op))
829 return PREP_ENCODED_CHECKSUM_ERR;
831 /* Last point we can still verify checksum: */
832 csum = bch2_checksum_bio(c, op->crc.csum_type,
833 extent_nonce(op->version, op->crc),
835 if (bch2_crc_cmp(op->crc.csum, csum))
836 return PREP_ENCODED_CHECKSUM_ERR;
838 if (bch2_bio_uncompress_inplace(c, bio, &op->crc))
839 return PREP_ENCODED_ERR;
843 * No longer have compressed data after this point - data might be
848 * If the data is checksummed and we're only writing a subset,
849 * rechecksum and adjust bio to point to currently live data:
851 if ((op->crc.live_size != op->crc.uncompressed_size ||
852 op->crc.csum_type != op->csum_type) &&
853 bch2_write_rechecksum(c, op, op->csum_type))
854 return PREP_ENCODED_CHECKSUM_ERR;
857 * If we want to compress the data, it has to be decrypted:
859 if ((op->compression_type ||
860 bch2_csum_type_is_encryption(op->crc.csum_type) !=
861 bch2_csum_type_is_encryption(op->csum_type)) &&
862 bch2_write_decrypt(op))
863 return PREP_ENCODED_CHECKSUM_ERR;
865 return PREP_ENCODED_OK;
868 static int bch2_write_extent(struct bch_write_op *op, struct write_point *wp,
871 struct bch_fs *c = op->c;
872 struct bio *src = &op->wbio.bio, *dst = src;
873 struct bvec_iter saved_iter;
875 unsigned total_output = 0, total_input = 0;
877 bool page_alloc_failed = false;
880 BUG_ON(!bio_sectors(src));
882 ec_buf = bch2_writepoint_ec_buf(c, wp);
884 switch (bch2_write_prep_encoded_data(op, wp)) {
885 case PREP_ENCODED_OK:
887 case PREP_ENCODED_ERR:
890 case PREP_ENCODED_CHECKSUM_ERR:
892 case PREP_ENCODED_DO_WRITE:
893 /* XXX look for bug here */
895 dst = bch2_write_bio_alloc(c, wp, src,
898 bio_copy_data(dst, src);
901 init_append_extent(op, wp, op->version, op->crc);
906 op->compression_type ||
908 !(op->flags & BCH_WRITE_PAGES_STABLE)) ||
909 (bch2_csum_type_is_encryption(op->csum_type) &&
910 !(op->flags & BCH_WRITE_PAGES_OWNED))) {
911 dst = bch2_write_bio_alloc(c, wp, src,
917 saved_iter = dst->bi_iter;
920 struct bch_extent_crc_unpacked crc =
921 (struct bch_extent_crc_unpacked) { 0 };
922 struct bversion version = op->version;
923 size_t dst_len, src_len;
925 if (page_alloc_failed &&
926 dst->bi_iter.bi_size < (wp->sectors_free << 9) &&
927 dst->bi_iter.bi_size < c->opts.encoded_extent_max)
930 BUG_ON(op->compression_type &&
931 (op->flags & BCH_WRITE_DATA_ENCODED) &&
932 bch2_csum_type_is_encryption(op->crc.csum_type));
933 BUG_ON(op->compression_type && !bounce);
935 crc.compression_type = op->incompressible
936 ? BCH_COMPRESSION_TYPE_incompressible
937 : op->compression_type
938 ? bch2_bio_compress(c, dst, &dst_len, src, &src_len,
939 op->compression_type)
941 if (!crc_is_compressed(crc)) {
942 dst_len = min(dst->bi_iter.bi_size, src->bi_iter.bi_size);
943 dst_len = min_t(unsigned, dst_len, wp->sectors_free << 9);
946 dst_len = min_t(unsigned, dst_len,
947 c->opts.encoded_extent_max);
950 swap(dst->bi_iter.bi_size, dst_len);
951 bio_copy_data(dst, src);
952 swap(dst->bi_iter.bi_size, dst_len);
958 BUG_ON(!src_len || !dst_len);
960 if (bch2_csum_type_is_encryption(op->csum_type)) {
961 if (bversion_zero(version)) {
962 version.lo = atomic64_inc_return(&c->key_version);
964 crc.nonce = op->nonce;
965 op->nonce += src_len >> 9;
969 if ((op->flags & BCH_WRITE_DATA_ENCODED) &&
970 !crc_is_compressed(crc) &&
971 bch2_csum_type_is_encryption(op->crc.csum_type) ==
972 bch2_csum_type_is_encryption(op->csum_type)) {
974 * Note: when we're using rechecksum(), we need to be
975 * checksumming @src because it has all the data our
976 * existing checksum covers - if we bounced (because we
977 * were trying to compress), @dst will only have the
978 * part of the data the new checksum will cover.
980 * But normally we want to be checksumming post bounce,
981 * because part of the reason for bouncing is so the
982 * data can't be modified (by userspace) while it's in
985 if (bch2_rechecksum_bio(c, src, version, op->crc,
988 bio_sectors(src) - (src_len >> 9),
992 if ((op->flags & BCH_WRITE_DATA_ENCODED) &&
993 bch2_rechecksum_bio(c, src, version, op->crc,
996 bio_sectors(src) - (src_len >> 9),
1000 crc.compressed_size = dst_len >> 9;
1001 crc.uncompressed_size = src_len >> 9;
1002 crc.live_size = src_len >> 9;
1004 swap(dst->bi_iter.bi_size, dst_len);
1005 ret = bch2_encrypt_bio(c, op->csum_type,
1006 extent_nonce(version, crc), dst);
1010 crc.csum = bch2_checksum_bio(c, op->csum_type,
1011 extent_nonce(version, crc), dst);
1012 crc.csum_type = op->csum_type;
1013 swap(dst->bi_iter.bi_size, dst_len);
1016 init_append_extent(op, wp, version, crc);
1019 bio_advance(dst, dst_len);
1020 bio_advance(src, src_len);
1021 total_output += dst_len;
1022 total_input += src_len;
1023 } while (dst->bi_iter.bi_size &&
1024 src->bi_iter.bi_size &&
1026 !bch2_keylist_realloc(&op->insert_keys,
1028 ARRAY_SIZE(op->inline_keys),
1029 BKEY_EXTENT_U64s_MAX));
1031 more = src->bi_iter.bi_size != 0;
1033 dst->bi_iter = saved_iter;
1035 if (dst == src && more) {
1036 BUG_ON(total_output != total_input);
1038 dst = bio_split(src, total_input >> 9,
1039 GFP_NOIO, &c->bio_write);
1040 wbio_init(dst)->put_bio = true;
1041 /* copy WRITE_SYNC flag */
1042 dst->bi_opf = src->bi_opf;
1045 dst->bi_iter.bi_size = total_output;
1050 bch_err(c, "error verifying existing checksum while rewriting existing data (memory corruption?)");
1053 if (to_wbio(dst)->bounce)
1054 bch2_bio_free_pages_pool(c, dst);
1055 if (to_wbio(dst)->put_bio)
1061 static void __bch2_write(struct closure *cl)
1063 struct bch_write_op *op = container_of(cl, struct bch_write_op, cl);
1064 struct bch_fs *c = op->c;
1065 struct write_point *wp;
1066 struct bio *bio = NULL;
1067 bool skip_put = true;
1068 unsigned nofs_flags;
1071 nofs_flags = memalloc_nofs_save();
1073 memset(&op->failed, 0, sizeof(op->failed));
1076 struct bkey_i *key_to_write;
1077 unsigned key_to_write_offset = op->insert_keys.top_p -
1078 op->insert_keys.keys_p;
1080 /* +1 for possible cache device: */
1081 if (op->open_buckets.nr + op->nr_replicas + 1 >
1082 ARRAY_SIZE(op->open_buckets.v))
1085 if (bch2_keylist_realloc(&op->insert_keys,
1087 ARRAY_SIZE(op->inline_keys),
1088 BKEY_EXTENT_U64s_MAX))
1092 * The copygc thread is now global, which means it's no longer
1093 * freeing up space on specific disks, which means that
1094 * allocations for specific disks may hang arbitrarily long:
1096 wp = bch2_alloc_sectors_start(c,
1098 op->opts.erasure_code && !(op->flags & BCH_WRITE_CACHED),
1102 op->nr_replicas_required,
1105 (op->flags & (BCH_WRITE_ALLOC_NOWAIT|
1106 BCH_WRITE_ONLY_SPECIFIED_DEVS)) ? NULL : cl);
1109 if (unlikely(IS_ERR(wp))) {
1110 if (unlikely(PTR_ERR(wp) != -EAGAIN)) {
1119 * It's possible for the allocator to fail, put us on the
1120 * freelist waitlist, and then succeed in one of various retry
1121 * paths: if that happens, we need to disable the skip_put
1122 * optimization because otherwise there won't necessarily be a
1123 * barrier before we free the bch_write_op:
1125 if (atomic_read(&cl->remaining) & CLOSURE_WAITING)
1128 bch2_open_bucket_get(c, wp, &op->open_buckets);
1129 ret = bch2_write_extent(op, wp, &bio);
1130 bch2_alloc_sectors_done(c, wp);
1139 * for the skip_put optimization this has to be set
1140 * before we submit the bio:
1142 op->flags |= BCH_WRITE_DONE;
1145 bio->bi_end_io = bch2_write_endio;
1146 bio->bi_private = &op->cl;
1147 bio->bi_opf |= REQ_OP_WRITE;
1150 closure_get(bio->bi_private);
1152 op->flags |= BCH_WRITE_SKIP_CLOSURE_PUT;
1154 key_to_write = (void *) (op->insert_keys.keys_p +
1155 key_to_write_offset);
1157 bch2_submit_wbio_replicas(to_wbio(bio), c, BCH_DATA_user,
1162 continue_at(cl, bch2_write_index, index_update_wq(op));
1164 memalloc_nofs_restore(nofs_flags);
1168 op->flags |= BCH_WRITE_DONE;
1170 continue_at(cl, bch2_write_index, index_update_wq(op));
1174 * If the write can't all be submitted at once, we generally want to
1175 * block synchronously as that signals backpressure to the caller.
1177 * However, if we're running out of a workqueue, we can't block here
1178 * because we'll be blocking other work items from completing:
1180 if (current->flags & PF_WQ_WORKER) {
1181 continue_at(cl, bch2_write_index, index_update_wq(op));
1187 if (!bch2_keylist_empty(&op->insert_keys)) {
1188 __bch2_write_index(op);
1191 op->flags |= BCH_WRITE_DONE;
1192 continue_at_nobarrier(cl, bch2_write_done, NULL);
1200 static void bch2_write_data_inline(struct bch_write_op *op, unsigned data_len)
1202 struct closure *cl = &op->cl;
1203 struct bio *bio = &op->wbio.bio;
1204 struct bvec_iter iter;
1205 struct bkey_i_inline_data *id;
1209 bch2_check_set_feature(op->c, BCH_FEATURE_inline_data);
1211 ret = bch2_keylist_realloc(&op->insert_keys, op->inline_keys,
1212 ARRAY_SIZE(op->inline_keys),
1213 BKEY_U64s + DIV_ROUND_UP(data_len, 8));
1219 sectors = bio_sectors(bio);
1220 op->pos.offset += sectors;
1222 id = bkey_inline_data_init(op->insert_keys.top);
1224 id->k.version = op->version;
1225 id->k.size = sectors;
1227 iter = bio->bi_iter;
1228 iter.bi_size = data_len;
1229 memcpy_from_bio(id->v.data, bio, iter);
1231 while (data_len & 7)
1232 id->v.data[data_len++] = '\0';
1233 set_bkey_val_bytes(&id->k, data_len);
1234 bch2_keylist_push(&op->insert_keys);
1236 op->flags |= BCH_WRITE_WROTE_DATA_INLINE;
1237 op->flags |= BCH_WRITE_DONE;
1239 continue_at_nobarrier(cl, bch2_write_index, NULL);
1242 bch2_write_done(&op->cl);
1246 * bch_write - handle a write to a cache device or flash only volume
1248 * This is the starting point for any data to end up in a cache device; it could
1249 * be from a normal write, or a writeback write, or a write to a flash only
1250 * volume - it's also used by the moving garbage collector to compact data in
1251 * mostly empty buckets.
1253 * It first writes the data to the cache, creating a list of keys to be inserted
1254 * (if the data won't fit in a single open bucket, there will be multiple keys);
1255 * after the data is written it calls bch_journal, and after the keys have been
1256 * added to the next journal write they're inserted into the btree.
1258 * If op->discard is true, instead of inserting the data it invalidates the
1259 * region of the cache represented by op->bio and op->inode.
1261 void bch2_write(struct closure *cl)
1263 struct bch_write_op *op = container_of(cl, struct bch_write_op, cl);
1264 struct bio *bio = &op->wbio.bio;
1265 struct bch_fs *c = op->c;
1268 BUG_ON(!op->nr_replicas);
1269 BUG_ON(!op->write_point.v);
1270 BUG_ON(!bkey_cmp(op->pos, POS_MAX));
1272 op->start_time = local_clock();
1273 bch2_keylist_init(&op->insert_keys, op->inline_keys);
1274 wbio_init(bio)->put_bio = false;
1276 if (bio->bi_iter.bi_size & (c->opts.block_size - 1)) {
1277 bch_err_inum_ratelimited(c, op->pos.inode,
1278 "misaligned write");
1283 if (c->opts.nochanges ||
1284 !percpu_ref_tryget_live(&c->writes)) {
1289 this_cpu_add(c->counters[BCH_COUNTER_io_write], bio_sectors(bio));
1290 bch2_increment_clock(c, bio_sectors(bio), WRITE);
1292 data_len = min_t(u64, bio->bi_iter.bi_size,
1293 op->new_i_size - (op->pos.offset << 9));
1295 if (c->opts.inline_data &&
1296 data_len <= min(block_bytes(c) / 2, 1024U)) {
1297 bch2_write_data_inline(op, data_len);
1301 continue_at_nobarrier(cl, __bch2_write, NULL);
1304 bch2_disk_reservation_put(c, &op->res);
1307 EBUG_ON(cl->parent);
1308 closure_debug_destroy(cl);
1315 /* Cache promotion on read */
1319 struct rcu_head rcu;
1322 struct rhash_head hash;
1325 struct data_update write;
1326 struct bio_vec bi_inline_vecs[0]; /* must be last */
1329 static const struct rhashtable_params bch_promote_params = {
1330 .head_offset = offsetof(struct promote_op, hash),
1331 .key_offset = offsetof(struct promote_op, pos),
1332 .key_len = sizeof(struct bpos),
1335 static inline bool should_promote(struct bch_fs *c, struct bkey_s_c k,
1337 struct bch_io_opts opts,
1340 if (!(flags & BCH_READ_MAY_PROMOTE))
1343 if (!opts.promote_target)
1346 if (bch2_bkey_has_target(c, k, opts.promote_target))
1349 if (bch2_target_congested(c, opts.promote_target)) {
1350 /* XXX trace this */
1354 if (rhashtable_lookup_fast(&c->promote_table, &pos,
1355 bch_promote_params))
1361 static void promote_free(struct bch_fs *c, struct promote_op *op)
1365 ret = rhashtable_remove_fast(&c->promote_table, &op->hash,
1366 bch_promote_params);
1368 percpu_ref_put(&c->writes);
1372 static void promote_done(struct closure *cl)
1374 struct promote_op *op =
1375 container_of(cl, struct promote_op, cl);
1376 struct bch_fs *c = op->write.op.c;
1378 bch2_time_stats_update(&c->times[BCH_TIME_data_promote],
1381 bch2_data_update_exit(&op->write);
1382 promote_free(c, op);
1385 static void promote_start(struct promote_op *op, struct bch_read_bio *rbio)
1387 struct closure *cl = &op->cl;
1388 struct bio *bio = &op->write.op.wbio.bio;
1390 trace_promote(&rbio->bio);
1392 /* we now own pages: */
1393 BUG_ON(!rbio->bounce);
1394 BUG_ON(rbio->bio.bi_vcnt > bio->bi_max_vecs);
1396 memcpy(bio->bi_io_vec, rbio->bio.bi_io_vec,
1397 sizeof(struct bio_vec) * rbio->bio.bi_vcnt);
1398 swap(bio->bi_vcnt, rbio->bio.bi_vcnt);
1400 closure_init(cl, NULL);
1401 bch2_data_update_read_done(&op->write, rbio->pick.crc, cl);
1402 closure_return_with_destructor(cl, promote_done);
1405 static struct promote_op *__promote_alloc(struct bch_fs *c,
1406 enum btree_id btree_id,
1409 struct extent_ptr_decoded *pick,
1410 struct bch_io_opts opts,
1412 struct bch_read_bio **rbio)
1414 struct promote_op *op = NULL;
1416 unsigned pages = DIV_ROUND_UP(sectors, PAGE_SECTORS);
1419 if (!percpu_ref_tryget_live(&c->writes))
1422 op = kzalloc(sizeof(*op) + sizeof(struct bio_vec) * pages, GFP_NOIO);
1426 op->start_time = local_clock();
1430 * We don't use the mempool here because extents that aren't
1431 * checksummed or compressed can be too big for the mempool:
1433 *rbio = kzalloc(sizeof(struct bch_read_bio) +
1434 sizeof(struct bio_vec) * pages,
1439 rbio_init(&(*rbio)->bio, opts);
1440 bio_init(&(*rbio)->bio, NULL, (*rbio)->bio.bi_inline_vecs, pages, 0);
1442 if (bch2_bio_alloc_pages(&(*rbio)->bio, sectors << 9,
1446 (*rbio)->bounce = true;
1447 (*rbio)->split = true;
1448 (*rbio)->kmalloc = true;
1450 if (rhashtable_lookup_insert_fast(&c->promote_table, &op->hash,
1451 bch_promote_params))
1454 bio = &op->write.op.wbio.bio;
1455 bio_init(bio, NULL, bio->bi_inline_vecs, pages, 0);
1457 ret = bch2_data_update_init(c, &op->write,
1458 writepoint_hashed((unsigned long) current),
1460 (struct data_update_opts) {
1461 .target = opts.promote_target,
1462 .extra_replicas = 1,
1463 .write_flags = BCH_WRITE_ALLOC_NOWAIT|BCH_WRITE_CACHED,
1471 bio_free_pages(&(*rbio)->bio);
1475 percpu_ref_put(&c->writes);
1480 static struct promote_op *promote_alloc(struct bch_fs *c,
1481 struct bvec_iter iter,
1483 struct extent_ptr_decoded *pick,
1484 struct bch_io_opts opts,
1486 struct bch_read_bio **rbio,
1490 bool promote_full = *read_full || READ_ONCE(c->promote_whole_extents);
1491 /* data might have to be decompressed in the write path: */
1492 unsigned sectors = promote_full
1493 ? max(pick->crc.compressed_size, pick->crc.live_size)
1494 : bvec_iter_sectors(iter);
1495 struct bpos pos = promote_full
1496 ? bkey_start_pos(k.k)
1497 : POS(k.k->p.inode, iter.bi_sector);
1498 struct promote_op *promote;
1500 if (!should_promote(c, k, pos, opts, flags))
1503 promote = __promote_alloc(c,
1504 k.k->type == KEY_TYPE_reflink_v
1507 k, pos, pick, opts, sectors, rbio);
1512 *read_full = promote_full;
1518 #define READ_RETRY_AVOID 1
1519 #define READ_RETRY 2
1524 RBIO_CONTEXT_HIGHPRI,
1525 RBIO_CONTEXT_UNBOUND,
1528 static inline struct bch_read_bio *
1529 bch2_rbio_parent(struct bch_read_bio *rbio)
1531 return rbio->split ? rbio->parent : rbio;
1535 static void bch2_rbio_punt(struct bch_read_bio *rbio, work_func_t fn,
1536 enum rbio_context context,
1537 struct workqueue_struct *wq)
1539 if (context <= rbio->context) {
1542 rbio->work.func = fn;
1543 rbio->context = context;
1544 queue_work(wq, &rbio->work);
1548 static inline struct bch_read_bio *bch2_rbio_free(struct bch_read_bio *rbio)
1550 BUG_ON(rbio->bounce && !rbio->split);
1553 promote_free(rbio->c, rbio->promote);
1554 rbio->promote = NULL;
1557 bch2_bio_free_pages_pool(rbio->c, &rbio->bio);
1560 struct bch_read_bio *parent = rbio->parent;
1565 bio_put(&rbio->bio);
1574 * Only called on a top level bch_read_bio to complete an entire read request,
1577 static void bch2_rbio_done(struct bch_read_bio *rbio)
1579 if (rbio->start_time)
1580 bch2_time_stats_update(&rbio->c->times[BCH_TIME_data_read],
1582 bio_endio(&rbio->bio);
1585 static void bch2_read_retry_nodecode(struct bch_fs *c, struct bch_read_bio *rbio,
1586 struct bvec_iter bvec_iter,
1587 struct bch_io_failures *failed,
1590 struct btree_trans trans;
1591 struct btree_iter iter;
1596 flags &= ~BCH_READ_LAST_FRAGMENT;
1597 flags |= BCH_READ_MUST_CLONE;
1599 bch2_bkey_buf_init(&sk);
1600 bch2_trans_init(&trans, c, 0, 0);
1602 bch2_trans_iter_init(&trans, &iter, rbio->data_btree,
1603 rbio->read_pos, BTREE_ITER_SLOTS);
1605 rbio->bio.bi_status = 0;
1607 k = bch2_btree_iter_peek_slot(&iter);
1611 bch2_bkey_buf_reassemble(&sk, c, k);
1612 k = bkey_i_to_s_c(sk.k);
1613 bch2_trans_unlock(&trans);
1615 if (!bch2_bkey_matches_ptr(c, k,
1617 rbio->data_pos.offset -
1618 rbio->pick.crc.offset)) {
1619 /* extent we wanted to read no longer exists: */
1624 ret = __bch2_read_extent(&trans, rbio, bvec_iter,
1627 k, 0, failed, flags);
1628 if (ret == READ_RETRY)
1633 bch2_rbio_done(rbio);
1634 bch2_trans_iter_exit(&trans, &iter);
1635 bch2_trans_exit(&trans);
1636 bch2_bkey_buf_exit(&sk, c);
1639 rbio->bio.bi_status = BLK_STS_IOERR;
1643 static void bch2_rbio_retry(struct work_struct *work)
1645 struct bch_read_bio *rbio =
1646 container_of(work, struct bch_read_bio, work);
1647 struct bch_fs *c = rbio->c;
1648 struct bvec_iter iter = rbio->bvec_iter;
1649 unsigned flags = rbio->flags;
1650 subvol_inum inum = {
1651 .subvol = rbio->subvol,
1652 .inum = rbio->read_pos.inode,
1654 struct bch_io_failures failed = { .nr = 0 };
1656 trace_read_retry(&rbio->bio);
1658 if (rbio->retry == READ_RETRY_AVOID)
1659 bch2_mark_io_failure(&failed, &rbio->pick);
1661 rbio->bio.bi_status = 0;
1663 rbio = bch2_rbio_free(rbio);
1665 flags |= BCH_READ_IN_RETRY;
1666 flags &= ~BCH_READ_MAY_PROMOTE;
1668 if (flags & BCH_READ_NODECODE) {
1669 bch2_read_retry_nodecode(c, rbio, iter, &failed, flags);
1671 flags &= ~BCH_READ_LAST_FRAGMENT;
1672 flags |= BCH_READ_MUST_CLONE;
1674 __bch2_read(c, rbio, iter, inum, &failed, flags);
1678 static void bch2_rbio_error(struct bch_read_bio *rbio, int retry,
1681 rbio->retry = retry;
1683 if (rbio->flags & BCH_READ_IN_RETRY)
1686 if (retry == READ_ERR) {
1687 rbio = bch2_rbio_free(rbio);
1689 rbio->bio.bi_status = error;
1690 bch2_rbio_done(rbio);
1692 bch2_rbio_punt(rbio, bch2_rbio_retry,
1693 RBIO_CONTEXT_UNBOUND, system_unbound_wq);
1697 static int __bch2_rbio_narrow_crcs(struct btree_trans *trans,
1698 struct bch_read_bio *rbio)
1700 struct bch_fs *c = rbio->c;
1701 u64 data_offset = rbio->data_pos.offset - rbio->pick.crc.offset;
1702 struct bch_extent_crc_unpacked new_crc;
1703 struct btree_iter iter;
1708 if (crc_is_compressed(rbio->pick.crc))
1711 bch2_trans_iter_init(trans, &iter, rbio->data_btree, rbio->data_pos,
1712 BTREE_ITER_SLOTS|BTREE_ITER_INTENT);
1713 k = bch2_btree_iter_peek_slot(&iter);
1714 if ((ret = bkey_err(k)))
1717 if (bversion_cmp(k.k->version, rbio->version) ||
1718 !bch2_bkey_matches_ptr(c, k, rbio->pick.ptr, data_offset))
1721 /* Extent was merged? */
1722 if (bkey_start_offset(k.k) < data_offset ||
1723 k.k->p.offset > data_offset + rbio->pick.crc.uncompressed_size)
1726 if (bch2_rechecksum_bio(c, &rbio->bio, rbio->version,
1727 rbio->pick.crc, NULL, &new_crc,
1728 bkey_start_offset(k.k) - data_offset, k.k->size,
1729 rbio->pick.crc.csum_type)) {
1730 bch_err(c, "error verifying existing checksum while narrowing checksum (memory corruption?)");
1736 * going to be temporarily appending another checksum entry:
1738 new = bch2_trans_kmalloc(trans, bkey_bytes(k.k) +
1739 sizeof(struct bch_extent_crc128));
1740 if ((ret = PTR_ERR_OR_ZERO(new)))
1743 bkey_reassemble(new, k);
1745 if (!bch2_bkey_narrow_crcs(new, new_crc))
1748 ret = bch2_trans_update(trans, &iter, new,
1749 BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE);
1751 bch2_trans_iter_exit(trans, &iter);
1755 static noinline void bch2_rbio_narrow_crcs(struct bch_read_bio *rbio)
1757 bch2_trans_do(rbio->c, NULL, NULL, BTREE_INSERT_NOFAIL,
1758 __bch2_rbio_narrow_crcs(&trans, rbio));
1761 /* Inner part that may run in process context */
1762 static void __bch2_read_endio(struct work_struct *work)
1764 struct bch_read_bio *rbio =
1765 container_of(work, struct bch_read_bio, work);
1766 struct bch_fs *c = rbio->c;
1767 struct bch_dev *ca = bch_dev_bkey_exists(c, rbio->pick.ptr.dev);
1768 struct bio *src = &rbio->bio;
1769 struct bio *dst = &bch2_rbio_parent(rbio)->bio;
1770 struct bvec_iter dst_iter = rbio->bvec_iter;
1771 struct bch_extent_crc_unpacked crc = rbio->pick.crc;
1772 struct nonce nonce = extent_nonce(rbio->version, crc);
1773 unsigned nofs_flags;
1774 struct bch_csum csum;
1777 nofs_flags = memalloc_nofs_save();
1779 /* Reset iterator for checksumming and copying bounced data: */
1781 src->bi_iter.bi_size = crc.compressed_size << 9;
1782 src->bi_iter.bi_idx = 0;
1783 src->bi_iter.bi_bvec_done = 0;
1785 src->bi_iter = rbio->bvec_iter;
1788 csum = bch2_checksum_bio(c, crc.csum_type, nonce, src);
1789 if (bch2_crc_cmp(csum, rbio->pick.crc.csum))
1794 * We need to rework the narrow_crcs path to deliver the read completion
1795 * first, and then punt to a different workqueue, otherwise we're
1796 * holding up reads while doing btree updates which is bad for memory
1799 if (unlikely(rbio->narrow_crcs))
1800 bch2_rbio_narrow_crcs(rbio);
1802 if (rbio->flags & BCH_READ_NODECODE)
1805 /* Adjust crc to point to subset of data we want: */
1806 crc.offset += rbio->offset_into_extent;
1807 crc.live_size = bvec_iter_sectors(rbio->bvec_iter);
1809 if (crc_is_compressed(crc)) {
1810 ret = bch2_encrypt_bio(c, crc.csum_type, nonce, src);
1814 if (bch2_bio_uncompress(c, src, dst, dst_iter, crc))
1815 goto decompression_err;
1817 /* don't need to decrypt the entire bio: */
1818 nonce = nonce_add(nonce, crc.offset << 9);
1819 bio_advance(src, crc.offset << 9);
1821 BUG_ON(src->bi_iter.bi_size < dst_iter.bi_size);
1822 src->bi_iter.bi_size = dst_iter.bi_size;
1824 ret = bch2_encrypt_bio(c, crc.csum_type, nonce, src);
1829 struct bvec_iter src_iter = src->bi_iter;
1830 bio_copy_data_iter(dst, &dst_iter, src, &src_iter);
1834 if (rbio->promote) {
1836 * Re encrypt data we decrypted, so it's consistent with
1839 ret = bch2_encrypt_bio(c, crc.csum_type, nonce, src);
1843 promote_start(rbio->promote, rbio);
1844 rbio->promote = NULL;
1847 if (likely(!(rbio->flags & BCH_READ_IN_RETRY))) {
1848 rbio = bch2_rbio_free(rbio);
1849 bch2_rbio_done(rbio);
1852 memalloc_nofs_restore(nofs_flags);
1856 * Checksum error: if the bio wasn't bounced, we may have been
1857 * reading into buffers owned by userspace (that userspace can
1858 * scribble over) - retry the read, bouncing it this time:
1860 if (!rbio->bounce && (rbio->flags & BCH_READ_USER_MAPPED)) {
1861 rbio->flags |= BCH_READ_MUST_BOUNCE;
1862 bch2_rbio_error(rbio, READ_RETRY, BLK_STS_IOERR);
1866 bch2_dev_inum_io_error(ca, rbio->read_pos.inode, (u64) rbio->bvec_iter.bi_sector,
1867 "data checksum error: expected %0llx:%0llx got %0llx:%0llx (type %s)",
1868 rbio->pick.crc.csum.hi, rbio->pick.crc.csum.lo,
1869 csum.hi, csum.lo, bch2_csum_types[crc.csum_type]);
1870 bch2_rbio_error(rbio, READ_RETRY_AVOID, BLK_STS_IOERR);
1873 bch_err_inum_ratelimited(c, rbio->read_pos.inode,
1874 "decompression error");
1875 bch2_rbio_error(rbio, READ_ERR, BLK_STS_IOERR);
1878 bch_err_inum_ratelimited(c, rbio->read_pos.inode,
1880 bch2_rbio_error(rbio, READ_ERR, BLK_STS_IOERR);
1884 static void bch2_read_endio(struct bio *bio)
1886 struct bch_read_bio *rbio =
1887 container_of(bio, struct bch_read_bio, bio);
1888 struct bch_fs *c = rbio->c;
1889 struct bch_dev *ca = bch_dev_bkey_exists(c, rbio->pick.ptr.dev);
1890 struct workqueue_struct *wq = NULL;
1891 enum rbio_context context = RBIO_CONTEXT_NULL;
1893 if (rbio->have_ioref) {
1894 bch2_latency_acct(ca, rbio->submit_time, READ);
1895 percpu_ref_put(&ca->io_ref);
1899 rbio->bio.bi_end_io = rbio->end_io;
1901 if (bch2_dev_inum_io_err_on(bio->bi_status, ca,
1902 rbio->read_pos.inode,
1903 rbio->read_pos.offset,
1904 "data read error: %s",
1905 bch2_blk_status_to_str(bio->bi_status))) {
1906 bch2_rbio_error(rbio, READ_RETRY_AVOID, bio->bi_status);
1910 if (((rbio->flags & BCH_READ_RETRY_IF_STALE) && race_fault()) ||
1911 ptr_stale(ca, &rbio->pick.ptr)) {
1912 atomic_long_inc(&c->read_realloc_races);
1914 if (rbio->flags & BCH_READ_RETRY_IF_STALE)
1915 bch2_rbio_error(rbio, READ_RETRY, BLK_STS_AGAIN);
1917 bch2_rbio_error(rbio, READ_ERR, BLK_STS_AGAIN);
1921 if (rbio->narrow_crcs ||
1923 crc_is_compressed(rbio->pick.crc) ||
1924 bch2_csum_type_is_encryption(rbio->pick.crc.csum_type))
1925 context = RBIO_CONTEXT_UNBOUND, wq = system_unbound_wq;
1926 else if (rbio->pick.crc.csum_type)
1927 context = RBIO_CONTEXT_HIGHPRI, wq = system_highpri_wq;
1929 bch2_rbio_punt(rbio, __bch2_read_endio, context, wq);
1932 int __bch2_read_indirect_extent(struct btree_trans *trans,
1933 unsigned *offset_into_extent,
1934 struct bkey_buf *orig_k)
1936 struct btree_iter iter;
1941 reflink_offset = le64_to_cpu(bkey_i_to_reflink_p(orig_k->k)->v.idx) +
1942 *offset_into_extent;
1944 bch2_trans_iter_init(trans, &iter, BTREE_ID_reflink,
1945 POS(0, reflink_offset),
1947 k = bch2_btree_iter_peek_slot(&iter);
1952 if (k.k->type != KEY_TYPE_reflink_v &&
1953 k.k->type != KEY_TYPE_indirect_inline_data) {
1954 bch_err_inum_ratelimited(trans->c, orig_k->k->k.p.inode,
1955 "%llu len %u points to nonexistent indirect extent %llu",
1956 orig_k->k->k.p.offset,
1959 bch2_inconsistent_error(trans->c);
1964 *offset_into_extent = iter.pos.offset - bkey_start_offset(k.k);
1965 bch2_bkey_buf_reassemble(orig_k, trans->c, k);
1967 bch2_trans_iter_exit(trans, &iter);
1971 static noinline void read_from_stale_dirty_pointer(struct btree_trans *trans,
1973 struct bch_extent_ptr ptr)
1975 struct bch_fs *c = trans->c;
1976 struct bch_dev *ca = bch_dev_bkey_exists(c, ptr.dev);
1977 struct btree_iter iter;
1978 struct printbuf buf = PRINTBUF;
1981 bch2_trans_iter_init(trans, &iter, BTREE_ID_alloc,
1982 PTR_BUCKET_POS(c, &ptr),
1985 prt_printf(&buf, "Attempting to read from stale dirty pointer:");
1986 printbuf_indent_add(&buf, 2);
1989 bch2_bkey_val_to_text(&buf, c, k);
1992 prt_printf(&buf, "memory gen: %u", *bucket_gen(ca, iter.pos.offset));
1994 ret = lockrestart_do(trans, bkey_err(k = bch2_btree_iter_peek_slot(&iter)));
1997 bch2_bkey_val_to_text(&buf, c, k);
2000 bch2_fs_inconsistent(c, "%s", buf.buf);
2002 bch2_trans_iter_exit(trans, &iter);
2003 printbuf_exit(&buf);
2006 int __bch2_read_extent(struct btree_trans *trans, struct bch_read_bio *orig,
2007 struct bvec_iter iter, struct bpos read_pos,
2008 enum btree_id data_btree, struct bkey_s_c k,
2009 unsigned offset_into_extent,
2010 struct bch_io_failures *failed, unsigned flags)
2012 struct bch_fs *c = trans->c;
2013 struct extent_ptr_decoded pick;
2014 struct bch_read_bio *rbio = NULL;
2015 struct bch_dev *ca = NULL;
2016 struct promote_op *promote = NULL;
2017 bool bounce = false, read_full = false, narrow_crcs = false;
2018 struct bpos data_pos = bkey_start_pos(k.k);
2021 if (bkey_extent_is_inline_data(k.k)) {
2022 unsigned bytes = min_t(unsigned, iter.bi_size,
2023 bkey_inline_data_bytes(k.k));
2025 swap(iter.bi_size, bytes);
2026 memcpy_to_bio(&orig->bio, iter, bkey_inline_data_p(k));
2027 swap(iter.bi_size, bytes);
2028 bio_advance_iter(&orig->bio, &iter, bytes);
2029 zero_fill_bio_iter(&orig->bio, iter);
2033 pick_ret = bch2_bkey_pick_read_device(c, k, failed, &pick);
2035 /* hole or reservation - just zero fill: */
2040 bch_err_inum_ratelimited(c, k.k->p.inode,
2041 "no device to read from");
2045 ca = bch_dev_bkey_exists(c, pick.ptr.dev);
2048 * Stale dirty pointers are treated as IO errors, but @failed isn't
2049 * allocated unless we're in the retry path - so if we're not in the
2050 * retry path, don't check here, it'll be caught in bch2_read_endio()
2051 * and we'll end up in the retry path:
2053 if ((flags & BCH_READ_IN_RETRY) &&
2055 unlikely(ptr_stale(ca, &pick.ptr))) {
2056 read_from_stale_dirty_pointer(trans, k, pick.ptr);
2057 bch2_mark_io_failure(failed, &pick);
2062 * Unlock the iterator while the btree node's lock is still in
2063 * cache, before doing the IO:
2065 bch2_trans_unlock(trans);
2067 if (flags & BCH_READ_NODECODE) {
2069 * can happen if we retry, and the extent we were going to read
2070 * has been merged in the meantime:
2072 if (pick.crc.compressed_size > orig->bio.bi_vcnt * PAGE_SECTORS)
2075 iter.bi_size = pick.crc.compressed_size << 9;
2079 if (!(flags & BCH_READ_LAST_FRAGMENT) ||
2080 bio_flagged(&orig->bio, BIO_CHAIN))
2081 flags |= BCH_READ_MUST_CLONE;
2083 narrow_crcs = !(flags & BCH_READ_IN_RETRY) &&
2084 bch2_can_narrow_extent_crcs(k, pick.crc);
2086 if (narrow_crcs && (flags & BCH_READ_USER_MAPPED))
2087 flags |= BCH_READ_MUST_BOUNCE;
2089 EBUG_ON(offset_into_extent + bvec_iter_sectors(iter) > k.k->size);
2091 if (crc_is_compressed(pick.crc) ||
2092 (pick.crc.csum_type != BCH_CSUM_none &&
2093 (bvec_iter_sectors(iter) != pick.crc.uncompressed_size ||
2094 (bch2_csum_type_is_encryption(pick.crc.csum_type) &&
2095 (flags & BCH_READ_USER_MAPPED)) ||
2096 (flags & BCH_READ_MUST_BOUNCE)))) {
2101 if (orig->opts.promote_target)
2102 promote = promote_alloc(c, iter, k, &pick, orig->opts, flags,
2103 &rbio, &bounce, &read_full);
2106 EBUG_ON(crc_is_compressed(pick.crc));
2107 EBUG_ON(pick.crc.csum_type &&
2108 (bvec_iter_sectors(iter) != pick.crc.uncompressed_size ||
2109 bvec_iter_sectors(iter) != pick.crc.live_size ||
2111 offset_into_extent));
2113 data_pos.offset += offset_into_extent;
2114 pick.ptr.offset += pick.crc.offset +
2116 offset_into_extent = 0;
2117 pick.crc.compressed_size = bvec_iter_sectors(iter);
2118 pick.crc.uncompressed_size = bvec_iter_sectors(iter);
2119 pick.crc.offset = 0;
2120 pick.crc.live_size = bvec_iter_sectors(iter);
2121 offset_into_extent = 0;
2126 * promote already allocated bounce rbio:
2127 * promote needs to allocate a bio big enough for uncompressing
2128 * data in the write path, but we're not going to use it all
2131 EBUG_ON(rbio->bio.bi_iter.bi_size <
2132 pick.crc.compressed_size << 9);
2133 rbio->bio.bi_iter.bi_size =
2134 pick.crc.compressed_size << 9;
2135 } else if (bounce) {
2136 unsigned sectors = pick.crc.compressed_size;
2138 rbio = rbio_init(bio_alloc_bioset(NULL,
2139 DIV_ROUND_UP(sectors, PAGE_SECTORS),
2142 &c->bio_read_split),
2145 bch2_bio_alloc_pages_pool(c, &rbio->bio, sectors << 9);
2146 rbio->bounce = true;
2148 } else if (flags & BCH_READ_MUST_CLONE) {
2150 * Have to clone if there were any splits, due to error
2151 * reporting issues (if a split errored, and retrying didn't
2152 * work, when it reports the error to its parent (us) we don't
2153 * know if the error was from our bio, and we should retry, or
2154 * from the whole bio, in which case we don't want to retry and
2157 rbio = rbio_init(bio_alloc_clone(NULL, &orig->bio, GFP_NOIO,
2158 &c->bio_read_split),
2160 rbio->bio.bi_iter = iter;
2164 rbio->bio.bi_iter = iter;
2165 EBUG_ON(bio_flagged(&rbio->bio, BIO_CHAIN));
2168 EBUG_ON(bio_sectors(&rbio->bio) != pick.crc.compressed_size);
2171 rbio->submit_time = local_clock();
2173 rbio->parent = orig;
2175 rbio->end_io = orig->bio.bi_end_io;
2176 rbio->bvec_iter = iter;
2177 rbio->offset_into_extent= offset_into_extent;
2178 rbio->flags = flags;
2179 rbio->have_ioref = pick_ret > 0 && bch2_dev_get_ioref(ca, READ);
2180 rbio->narrow_crcs = narrow_crcs;
2184 /* XXX: only initialize this if needed */
2185 rbio->devs_have = bch2_bkey_devs(k);
2187 rbio->subvol = orig->subvol;
2188 rbio->read_pos = read_pos;
2189 rbio->data_btree = data_btree;
2190 rbio->data_pos = data_pos;
2191 rbio->version = k.k->version;
2192 rbio->promote = promote;
2193 INIT_WORK(&rbio->work, NULL);
2195 rbio->bio.bi_opf = orig->bio.bi_opf;
2196 rbio->bio.bi_iter.bi_sector = pick.ptr.offset;
2197 rbio->bio.bi_end_io = bch2_read_endio;
2200 trace_read_bounce(&rbio->bio);
2202 this_cpu_add(c->counters[BCH_COUNTER_io_read], bio_sectors(&rbio->bio));
2203 bch2_increment_clock(c, bio_sectors(&rbio->bio), READ);
2206 * If it's being moved internally, we don't want to flag it as a cache
2209 if (pick.ptr.cached && !(flags & BCH_READ_NODECODE))
2210 bch2_bucket_io_time_reset(trans, pick.ptr.dev,
2211 PTR_BUCKET_NR(ca, &pick.ptr), READ);
2213 if (!(flags & (BCH_READ_IN_RETRY|BCH_READ_LAST_FRAGMENT))) {
2214 bio_inc_remaining(&orig->bio);
2215 trace_read_split(&orig->bio);
2218 if (!rbio->pick.idx) {
2219 if (!rbio->have_ioref) {
2220 bch_err_inum_ratelimited(c, k.k->p.inode,
2221 "no device to read from");
2222 bch2_rbio_error(rbio, READ_RETRY_AVOID, BLK_STS_IOERR);
2226 this_cpu_add(ca->io_done->sectors[READ][BCH_DATA_user],
2227 bio_sectors(&rbio->bio));
2228 bio_set_dev(&rbio->bio, ca->disk_sb.bdev);
2230 if (likely(!(flags & BCH_READ_IN_RETRY)))
2231 submit_bio(&rbio->bio);
2233 submit_bio_wait(&rbio->bio);
2235 /* Attempting reconstruct read: */
2236 if (bch2_ec_read_extent(c, rbio)) {
2237 bch2_rbio_error(rbio, READ_RETRY_AVOID, BLK_STS_IOERR);
2241 if (likely(!(flags & BCH_READ_IN_RETRY)))
2242 bio_endio(&rbio->bio);
2245 if (likely(!(flags & BCH_READ_IN_RETRY))) {
2250 rbio->context = RBIO_CONTEXT_UNBOUND;
2251 bch2_read_endio(&rbio->bio);
2254 rbio = bch2_rbio_free(rbio);
2256 if (ret == READ_RETRY_AVOID) {
2257 bch2_mark_io_failure(failed, &pick);
2268 if (flags & BCH_READ_IN_RETRY)
2271 orig->bio.bi_status = BLK_STS_IOERR;
2276 * won't normally happen in the BCH_READ_NODECODE
2277 * (bch2_move_extent()) path, but if we retry and the extent we wanted
2278 * to read no longer exists we have to signal that:
2280 if (flags & BCH_READ_NODECODE)
2283 zero_fill_bio_iter(&orig->bio, iter);
2285 if (flags & BCH_READ_LAST_FRAGMENT)
2286 bch2_rbio_done(orig);
2290 void __bch2_read(struct bch_fs *c, struct bch_read_bio *rbio,
2291 struct bvec_iter bvec_iter, subvol_inum inum,
2292 struct bch_io_failures *failed, unsigned flags)
2294 struct btree_trans trans;
2295 struct btree_iter iter;
2301 BUG_ON(flags & BCH_READ_NODECODE);
2303 bch2_bkey_buf_init(&sk);
2304 bch2_trans_init(&trans, c, 0, 0);
2306 bch2_trans_begin(&trans);
2307 iter = (struct btree_iter) { NULL };
2309 ret = bch2_subvolume_get_snapshot(&trans, inum.subvol, &snapshot);
2313 bch2_trans_iter_init(&trans, &iter, BTREE_ID_extents,
2314 SPOS(inum.inum, bvec_iter.bi_sector, snapshot),
2317 unsigned bytes, sectors, offset_into_extent;
2318 enum btree_id data_btree = BTREE_ID_extents;
2321 * read_extent -> io_time_reset may cause a transaction restart
2322 * without returning an error, we need to check for that here:
2324 ret = bch2_trans_relock(&trans);
2328 bch2_btree_iter_set_pos(&iter,
2329 POS(inum.inum, bvec_iter.bi_sector));
2331 k = bch2_btree_iter_peek_slot(&iter);
2336 offset_into_extent = iter.pos.offset -
2337 bkey_start_offset(k.k);
2338 sectors = k.k->size - offset_into_extent;
2340 bch2_bkey_buf_reassemble(&sk, c, k);
2342 ret = bch2_read_indirect_extent(&trans, &data_btree,
2343 &offset_into_extent, &sk);
2347 k = bkey_i_to_s_c(sk.k);
2350 * With indirect extents, the amount of data to read is the min
2351 * of the original extent and the indirect extent:
2353 sectors = min(sectors, k.k->size - offset_into_extent);
2355 bytes = min(sectors, bvec_iter_sectors(bvec_iter)) << 9;
2356 swap(bvec_iter.bi_size, bytes);
2358 if (bvec_iter.bi_size == bytes)
2359 flags |= BCH_READ_LAST_FRAGMENT;
2361 ret = __bch2_read_extent(&trans, rbio, bvec_iter, iter.pos,
2363 offset_into_extent, failed, flags);
2367 if (flags & BCH_READ_LAST_FRAGMENT)
2370 swap(bvec_iter.bi_size, bytes);
2371 bio_advance_iter(&rbio->bio, &bvec_iter, bytes);
2373 ret = btree_trans_too_many_iters(&trans);
2378 bch2_trans_iter_exit(&trans, &iter);
2380 if (bch2_err_matches(ret, BCH_ERR_transaction_restart) ||
2381 ret == READ_RETRY ||
2382 ret == READ_RETRY_AVOID)
2385 bch2_trans_exit(&trans);
2386 bch2_bkey_buf_exit(&sk, c);
2389 bch_err_inum_ratelimited(c, inum.inum,
2390 "read error %i from btree lookup", ret);
2391 rbio->bio.bi_status = BLK_STS_IOERR;
2392 bch2_rbio_done(rbio);
2396 void bch2_fs_io_exit(struct bch_fs *c)
2398 if (c->promote_table.tbl)
2399 rhashtable_destroy(&c->promote_table);
2400 mempool_exit(&c->bio_bounce_pages);
2401 bioset_exit(&c->bio_write);
2402 bioset_exit(&c->bio_read_split);
2403 bioset_exit(&c->bio_read);
2406 int bch2_fs_io_init(struct bch_fs *c)
2408 if (bioset_init(&c->bio_read, 1, offsetof(struct bch_read_bio, bio),
2409 BIOSET_NEED_BVECS) ||
2410 bioset_init(&c->bio_read_split, 1, offsetof(struct bch_read_bio, bio),
2411 BIOSET_NEED_BVECS) ||
2412 bioset_init(&c->bio_write, 1, offsetof(struct bch_write_bio, bio),
2413 BIOSET_NEED_BVECS) ||
2414 mempool_init_page_pool(&c->bio_bounce_pages,
2416 c->opts.btree_node_size,
2417 c->opts.encoded_extent_max) /
2419 rhashtable_init(&c->promote_table, &bch_promote_params))