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 -EINTR 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;
328 while (!ret || ret == -EINTR) {
329 struct disk_reservation disk_res =
330 bch2_disk_reservation_init(c, 0);
331 struct bkey_i delete;
336 bch2_trans_begin(trans);
338 ret = bch2_subvolume_get_snapshot(trans, inum.subvol, &snapshot);
342 bch2_btree_iter_set_snapshot(iter, snapshot);
344 k = bch2_btree_iter_peek(iter);
345 if (bkey_cmp(iter->pos, end_pos) >= 0) {
346 bch2_btree_iter_set_pos(iter, end_pos);
354 bkey_init(&delete.k);
355 delete.k.p = iter->pos;
357 /* create the biggest key we can */
358 bch2_key_resize(&delete.k, max_sectors);
359 bch2_cut_back(end_pos, &delete);
361 ret = bch2_extent_update(trans, inum, iter, &delete,
363 0, i_sectors_delta, false);
364 bch2_disk_reservation_put(c, &disk_res);
370 int bch2_fpunch(struct bch_fs *c, subvol_inum inum, u64 start, u64 end,
371 s64 *i_sectors_delta)
373 struct btree_trans trans;
374 struct btree_iter iter;
377 bch2_trans_init(&trans, c, BTREE_ITER_MAX, 1024);
378 bch2_trans_iter_init(&trans, &iter, BTREE_ID_extents,
379 POS(inum.inum, start),
382 ret = bch2_fpunch_at(&trans, &iter, inum, end, i_sectors_delta);
384 bch2_trans_iter_exit(&trans, &iter);
385 bch2_trans_exit(&trans);
387 return ret == -EINTR ? 0 : ret;
390 int bch2_write_index_default(struct bch_write_op *op)
392 struct bch_fs *c = op->c;
394 struct open_bucket *ec_ob = ec_open_bucket(c, &op->open_buckets);
395 struct keylist *keys = &op->insert_keys;
396 struct bkey_i *k = bch2_keylist_front(keys);
397 struct btree_trans trans;
398 struct btree_iter iter;
400 .subvol = op->subvol,
401 .inum = k->k.p.inode,
405 BUG_ON(!inum.subvol);
407 bch2_bkey_buf_init(&sk);
408 bch2_trans_init(&trans, c, BTREE_ITER_MAX, 1024);
411 bch2_trans_begin(&trans);
413 k = bch2_keylist_front(keys);
414 bch2_bkey_buf_copy(&sk, c, k);
416 ret = bch2_subvolume_get_snapshot(&trans, inum.subvol,
417 &sk.k->k.p.snapshot);
423 bch2_trans_iter_init(&trans, &iter, BTREE_ID_extents,
424 bkey_start_pos(&sk.k->k),
425 BTREE_ITER_SLOTS|BTREE_ITER_INTENT);
427 ret = bch2_extent_update(&trans, inum, &iter, sk.k,
428 &op->res, op_journal_seq(op),
429 op->new_i_size, &op->i_sectors_delta,
430 op->flags & BCH_WRITE_CHECK_ENOSPC);
431 bch2_trans_iter_exit(&trans, &iter);
439 bch2_ob_add_backpointer(c, ec_ob, &sk.k->k);
441 if (bkey_cmp(iter.pos, k->k.p) >= 0)
442 bch2_keylist_pop_front(&op->insert_keys);
444 bch2_cut_front(iter.pos, k);
445 } while (!bch2_keylist_empty(keys));
447 bch2_trans_exit(&trans);
448 bch2_bkey_buf_exit(&sk, c);
455 void bch2_submit_wbio_replicas(struct bch_write_bio *wbio, struct bch_fs *c,
456 enum bch_data_type type,
457 const struct bkey_i *k)
459 struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(bkey_i_to_s_c(k));
460 const struct bch_extent_ptr *ptr;
461 struct bch_write_bio *n;
464 BUG_ON(c->opts.nochanges);
466 bkey_for_each_ptr(ptrs, ptr) {
467 BUG_ON(ptr->dev >= BCH_SB_MEMBERS_MAX ||
470 ca = bch_dev_bkey_exists(c, ptr->dev);
472 if (to_entry(ptr + 1) < ptrs.end) {
473 n = to_wbio(bio_clone_fast(&wbio->bio, GFP_NOIO,
476 n->bio.bi_end_io = wbio->bio.bi_end_io;
477 n->bio.bi_private = wbio->bio.bi_private;
482 n->bio.bi_opf = wbio->bio.bi_opf;
483 bio_inc_remaining(&wbio->bio);
491 n->have_ioref = bch2_dev_get_ioref(ca,
492 type == BCH_DATA_btree ? READ : WRITE);
493 n->submit_time = local_clock();
494 n->bio.bi_iter.bi_sector = ptr->offset;
496 if (likely(n->have_ioref)) {
497 this_cpu_add(ca->io_done->sectors[WRITE][type],
498 bio_sectors(&n->bio));
500 bio_set_dev(&n->bio, ca->disk_sb.bdev);
503 n->bio.bi_status = BLK_STS_REMOVED;
509 static void __bch2_write(struct closure *);
511 static void bch2_write_done(struct closure *cl)
513 struct bch_write_op *op = container_of(cl, struct bch_write_op, cl);
514 struct bch_fs *c = op->c;
516 if (!op->error && (op->flags & BCH_WRITE_FLUSH))
517 op->error = bch2_journal_error(&c->journal);
519 bch2_disk_reservation_put(c, &op->res);
520 percpu_ref_put(&c->writes);
521 bch2_keylist_free(&op->insert_keys, op->inline_keys);
523 bch2_time_stats_update(&c->times[BCH_TIME_data_write], op->start_time);
527 closure_debug_destroy(cl);
535 * bch_write_index - after a write, update index to point to new data
537 static void __bch2_write_index(struct bch_write_op *op)
539 struct bch_fs *c = op->c;
540 struct keylist *keys = &op->insert_keys;
541 struct bch_extent_ptr *ptr;
542 struct bkey_i *src, *dst = keys->keys, *n, *k;
546 for (src = keys->keys; src != keys->top; src = n) {
549 if (bkey_extent_is_direct_data(&src->k)) {
550 bch2_bkey_drop_ptrs(bkey_i_to_s(src), ptr,
551 test_bit(ptr->dev, op->failed.d));
553 if (!bch2_bkey_nr_ptrs(bkey_i_to_s_c(src))) {
560 memmove_u64s_down(dst, src, src->u64s);
561 dst = bkey_next(dst);
567 * probably not the ideal place to hook this in, but I don't
568 * particularly want to plumb io_opts all the way through the btree
569 * update stack right now
571 for_each_keylist_key(keys, k) {
572 bch2_rebalance_add_key(c, bkey_i_to_s_c(k), &op->opts);
574 if (bch2_bkey_is_incompressible(bkey_i_to_s_c(k)))
575 bch2_check_set_feature(op->c, BCH_FEATURE_incompressible);
579 if (!bch2_keylist_empty(keys)) {
580 u64 sectors_start = keylist_sectors(keys);
581 int ret = op->index_update_fn(op);
583 BUG_ON(ret == -EINTR);
584 BUG_ON(keylist_sectors(keys) && !ret);
586 op->written += sectors_start - keylist_sectors(keys);
589 bch_err_inum_ratelimited(c, op->pos.inode,
590 "write error %i from btree update", ret);
595 /* If some a bucket wasn't written, we can't erasure code it: */
596 for_each_set_bit(dev, op->failed.d, BCH_SB_MEMBERS_MAX)
597 bch2_open_bucket_write_error(c, &op->open_buckets, dev);
599 bch2_open_buckets_put(c, &op->open_buckets);
602 keys->top = keys->keys;
607 static void bch2_write_index(struct closure *cl)
609 struct bch_write_op *op = container_of(cl, struct bch_write_op, cl);
610 struct bch_fs *c = op->c;
612 __bch2_write_index(op);
614 if (!(op->flags & BCH_WRITE_DONE)) {
615 continue_at(cl, __bch2_write, index_update_wq(op));
616 } else if (!op->error && (op->flags & BCH_WRITE_FLUSH)) {
617 bch2_journal_flush_seq_async(&c->journal,
620 continue_at(cl, bch2_write_done, index_update_wq(op));
622 continue_at_nobarrier(cl, bch2_write_done, NULL);
626 static void bch2_write_endio(struct bio *bio)
628 struct closure *cl = bio->bi_private;
629 struct bch_write_op *op = container_of(cl, struct bch_write_op, cl);
630 struct bch_write_bio *wbio = to_wbio(bio);
631 struct bch_write_bio *parent = wbio->split ? wbio->parent : NULL;
632 struct bch_fs *c = wbio->c;
633 struct bch_dev *ca = bch_dev_bkey_exists(c, wbio->dev);
635 if (bch2_dev_inum_io_err_on(bio->bi_status, ca,
637 op->pos.offset - bio_sectors(bio), /* XXX definitely wrong */
638 "data write error: %s",
639 bch2_blk_status_to_str(bio->bi_status)))
640 set_bit(wbio->dev, op->failed.d);
642 if (wbio->have_ioref) {
643 bch2_latency_acct(ca, wbio->submit_time, WRITE);
644 percpu_ref_put(&ca->io_ref);
648 bch2_bio_free_pages_pool(c, bio);
654 bio_endio(&parent->bio);
655 else if (!(op->flags & BCH_WRITE_SKIP_CLOSURE_PUT))
658 continue_at_nobarrier(cl, bch2_write_index, index_update_wq(op));
661 static void init_append_extent(struct bch_write_op *op,
662 struct write_point *wp,
663 struct bversion version,
664 struct bch_extent_crc_unpacked crc)
666 struct bch_fs *c = op->c;
667 struct bkey_i_extent *e;
669 op->pos.offset += crc.uncompressed_size;
671 e = bkey_extent_init(op->insert_keys.top);
673 e->k.size = crc.uncompressed_size;
674 e->k.version = version;
677 crc.compression_type ||
679 bch2_extent_crc_append(&e->k_i, crc);
681 bch2_alloc_sectors_append_ptrs(c, wp, &e->k_i, crc.compressed_size,
682 op->flags & BCH_WRITE_CACHED);
684 bch2_keylist_push(&op->insert_keys);
687 static struct bio *bch2_write_bio_alloc(struct bch_fs *c,
688 struct write_point *wp,
690 bool *page_alloc_failed,
693 struct bch_write_bio *wbio;
695 unsigned output_available =
696 min(wp->sectors_free << 9, src->bi_iter.bi_size);
697 unsigned pages = DIV_ROUND_UP(output_available +
699 ? ((unsigned long) buf & (PAGE_SIZE - 1))
702 pages = min(pages, BIO_MAX_VECS);
704 bio = bio_alloc_bioset(GFP_NOIO, pages, &c->bio_write);
705 wbio = wbio_init(bio);
706 wbio->put_bio = true;
707 /* copy WRITE_SYNC flag */
708 wbio->bio.bi_opf = src->bi_opf;
711 bch2_bio_map(bio, buf, output_available);
718 * We can't use mempool for more than c->sb.encoded_extent_max
719 * worth of pages, but we'd like to allocate more if we can:
721 bch2_bio_alloc_pages_pool(c, bio,
722 min_t(unsigned, output_available,
723 c->opts.encoded_extent_max));
725 if (bio->bi_iter.bi_size < output_available)
727 bch2_bio_alloc_pages(bio,
729 bio->bi_iter.bi_size,
735 static int bch2_write_rechecksum(struct bch_fs *c,
736 struct bch_write_op *op,
737 unsigned new_csum_type)
739 struct bio *bio = &op->wbio.bio;
740 struct bch_extent_crc_unpacked new_crc;
743 /* bch2_rechecksum_bio() can't encrypt or decrypt data: */
745 if (bch2_csum_type_is_encryption(op->crc.csum_type) !=
746 bch2_csum_type_is_encryption(new_csum_type))
747 new_csum_type = op->crc.csum_type;
749 ret = bch2_rechecksum_bio(c, bio, op->version, op->crc,
751 op->crc.offset, op->crc.live_size,
756 bio_advance(bio, op->crc.offset << 9);
757 bio->bi_iter.bi_size = op->crc.live_size << 9;
762 static int bch2_write_decrypt(struct bch_write_op *op)
764 struct bch_fs *c = op->c;
765 struct nonce nonce = extent_nonce(op->version, op->crc);
766 struct bch_csum csum;
769 if (!bch2_csum_type_is_encryption(op->crc.csum_type))
773 * If we need to decrypt data in the write path, we'll no longer be able
774 * to verify the existing checksum (poly1305 mac, in this case) after
775 * it's decrypted - this is the last point we'll be able to reverify the
778 csum = bch2_checksum_bio(c, op->crc.csum_type, nonce, &op->wbio.bio);
779 if (bch2_crc_cmp(op->crc.csum, csum))
782 ret = bch2_encrypt_bio(c, op->crc.csum_type, nonce, &op->wbio.bio);
783 op->crc.csum_type = 0;
784 op->crc.csum = (struct bch_csum) { 0, 0 };
788 static enum prep_encoded_ret {
791 PREP_ENCODED_CHECKSUM_ERR,
792 PREP_ENCODED_DO_WRITE,
793 } bch2_write_prep_encoded_data(struct bch_write_op *op, struct write_point *wp)
795 struct bch_fs *c = op->c;
796 struct bio *bio = &op->wbio.bio;
798 if (!(op->flags & BCH_WRITE_DATA_ENCODED))
799 return PREP_ENCODED_OK;
801 BUG_ON(bio_sectors(bio) != op->crc.compressed_size);
803 /* Can we just write the entire extent as is? */
804 if (op->crc.uncompressed_size == op->crc.live_size &&
805 op->crc.compressed_size <= wp->sectors_free &&
806 (op->crc.compression_type == op->compression_type ||
807 op->incompressible)) {
808 if (!crc_is_compressed(op->crc) &&
809 op->csum_type != op->crc.csum_type &&
810 bch2_write_rechecksum(c, op, op->csum_type))
811 return PREP_ENCODED_CHECKSUM_ERR;
813 return PREP_ENCODED_DO_WRITE;
817 * If the data is compressed and we couldn't write the entire extent as
818 * is, we have to decompress it:
820 if (crc_is_compressed(op->crc)) {
821 struct bch_csum csum;
823 if (bch2_write_decrypt(op))
824 return PREP_ENCODED_CHECKSUM_ERR;
826 /* Last point we can still verify checksum: */
827 csum = bch2_checksum_bio(c, op->crc.csum_type,
828 extent_nonce(op->version, op->crc),
830 if (bch2_crc_cmp(op->crc.csum, csum))
831 return PREP_ENCODED_CHECKSUM_ERR;
833 if (bch2_bio_uncompress_inplace(c, bio, &op->crc))
834 return PREP_ENCODED_ERR;
838 * No longer have compressed data after this point - data might be
843 * If the data is checksummed and we're only writing a subset,
844 * rechecksum and adjust bio to point to currently live data:
846 if ((op->crc.live_size != op->crc.uncompressed_size ||
847 op->crc.csum_type != op->csum_type) &&
848 bch2_write_rechecksum(c, op, op->csum_type))
849 return PREP_ENCODED_CHECKSUM_ERR;
852 * If we want to compress the data, it has to be decrypted:
854 if ((op->compression_type ||
855 bch2_csum_type_is_encryption(op->crc.csum_type) !=
856 bch2_csum_type_is_encryption(op->csum_type)) &&
857 bch2_write_decrypt(op))
858 return PREP_ENCODED_CHECKSUM_ERR;
860 return PREP_ENCODED_OK;
863 static int bch2_write_extent(struct bch_write_op *op, struct write_point *wp,
866 struct bch_fs *c = op->c;
867 struct bio *src = &op->wbio.bio, *dst = src;
868 struct bvec_iter saved_iter;
870 unsigned total_output = 0, total_input = 0;
872 bool page_alloc_failed = false;
875 BUG_ON(!bio_sectors(src));
877 ec_buf = bch2_writepoint_ec_buf(c, wp);
879 switch (bch2_write_prep_encoded_data(op, wp)) {
880 case PREP_ENCODED_OK:
882 case PREP_ENCODED_ERR:
885 case PREP_ENCODED_CHECKSUM_ERR:
887 case PREP_ENCODED_DO_WRITE:
888 /* XXX look for bug here */
890 dst = bch2_write_bio_alloc(c, wp, src,
893 bio_copy_data(dst, src);
896 init_append_extent(op, wp, op->version, op->crc);
901 op->compression_type ||
903 !(op->flags & BCH_WRITE_PAGES_STABLE)) ||
904 (bch2_csum_type_is_encryption(op->csum_type) &&
905 !(op->flags & BCH_WRITE_PAGES_OWNED))) {
906 dst = bch2_write_bio_alloc(c, wp, src,
912 saved_iter = dst->bi_iter;
915 struct bch_extent_crc_unpacked crc =
916 (struct bch_extent_crc_unpacked) { 0 };
917 struct bversion version = op->version;
918 size_t dst_len, src_len;
920 if (page_alloc_failed &&
921 dst->bi_iter.bi_size < (wp->sectors_free << 9) &&
922 dst->bi_iter.bi_size < c->opts.encoded_extent_max)
925 BUG_ON(op->compression_type &&
926 (op->flags & BCH_WRITE_DATA_ENCODED) &&
927 bch2_csum_type_is_encryption(op->crc.csum_type));
928 BUG_ON(op->compression_type && !bounce);
930 crc.compression_type = op->incompressible
931 ? BCH_COMPRESSION_TYPE_incompressible
932 : op->compression_type
933 ? bch2_bio_compress(c, dst, &dst_len, src, &src_len,
934 op->compression_type)
936 if (!crc_is_compressed(crc)) {
937 dst_len = min(dst->bi_iter.bi_size, src->bi_iter.bi_size);
938 dst_len = min_t(unsigned, dst_len, wp->sectors_free << 9);
941 dst_len = min_t(unsigned, dst_len,
942 c->opts.encoded_extent_max);
945 swap(dst->bi_iter.bi_size, dst_len);
946 bio_copy_data(dst, src);
947 swap(dst->bi_iter.bi_size, dst_len);
953 BUG_ON(!src_len || !dst_len);
955 if (bch2_csum_type_is_encryption(op->csum_type)) {
956 if (bversion_zero(version)) {
957 version.lo = atomic64_inc_return(&c->key_version);
959 crc.nonce = op->nonce;
960 op->nonce += src_len >> 9;
964 if ((op->flags & BCH_WRITE_DATA_ENCODED) &&
965 !crc_is_compressed(crc) &&
966 bch2_csum_type_is_encryption(op->crc.csum_type) ==
967 bch2_csum_type_is_encryption(op->csum_type)) {
969 * Note: when we're using rechecksum(), we need to be
970 * checksumming @src because it has all the data our
971 * existing checksum covers - if we bounced (because we
972 * were trying to compress), @dst will only have the
973 * part of the data the new checksum will cover.
975 * But normally we want to be checksumming post bounce,
976 * because part of the reason for bouncing is so the
977 * data can't be modified (by userspace) while it's in
980 if (bch2_rechecksum_bio(c, src, version, op->crc,
983 bio_sectors(src) - (src_len >> 9),
987 if ((op->flags & BCH_WRITE_DATA_ENCODED) &&
988 bch2_rechecksum_bio(c, src, version, op->crc,
991 bio_sectors(src) - (src_len >> 9),
995 crc.compressed_size = dst_len >> 9;
996 crc.uncompressed_size = src_len >> 9;
997 crc.live_size = src_len >> 9;
999 swap(dst->bi_iter.bi_size, dst_len);
1000 ret = bch2_encrypt_bio(c, op->csum_type,
1001 extent_nonce(version, crc), dst);
1005 crc.csum = bch2_checksum_bio(c, op->csum_type,
1006 extent_nonce(version, crc), dst);
1007 crc.csum_type = op->csum_type;
1008 swap(dst->bi_iter.bi_size, dst_len);
1011 init_append_extent(op, wp, version, crc);
1014 bio_advance(dst, dst_len);
1015 bio_advance(src, src_len);
1016 total_output += dst_len;
1017 total_input += src_len;
1018 } while (dst->bi_iter.bi_size &&
1019 src->bi_iter.bi_size &&
1021 !bch2_keylist_realloc(&op->insert_keys,
1023 ARRAY_SIZE(op->inline_keys),
1024 BKEY_EXTENT_U64s_MAX));
1026 more = src->bi_iter.bi_size != 0;
1028 dst->bi_iter = saved_iter;
1030 if (dst == src && more) {
1031 BUG_ON(total_output != total_input);
1033 dst = bio_split(src, total_input >> 9,
1034 GFP_NOIO, &c->bio_write);
1035 wbio_init(dst)->put_bio = true;
1036 /* copy WRITE_SYNC flag */
1037 dst->bi_opf = src->bi_opf;
1040 dst->bi_iter.bi_size = total_output;
1045 bch_err(c, "error verifying existing checksum while "
1046 "rewriting existing data (memory corruption?)");
1049 if (to_wbio(dst)->bounce)
1050 bch2_bio_free_pages_pool(c, dst);
1051 if (to_wbio(dst)->put_bio)
1057 static void __bch2_write(struct closure *cl)
1059 struct bch_write_op *op = container_of(cl, struct bch_write_op, cl);
1060 struct bch_fs *c = op->c;
1061 struct write_point *wp;
1062 struct bio *bio = NULL;
1063 bool skip_put = true;
1064 unsigned nofs_flags;
1067 nofs_flags = memalloc_nofs_save();
1069 memset(&op->failed, 0, sizeof(op->failed));
1072 struct bkey_i *key_to_write;
1073 unsigned key_to_write_offset = op->insert_keys.top_p -
1074 op->insert_keys.keys_p;
1076 /* +1 for possible cache device: */
1077 if (op->open_buckets.nr + op->nr_replicas + 1 >
1078 ARRAY_SIZE(op->open_buckets.v))
1081 if (bch2_keylist_realloc(&op->insert_keys,
1083 ARRAY_SIZE(op->inline_keys),
1084 BKEY_EXTENT_U64s_MAX))
1087 if ((op->flags & BCH_WRITE_FROM_INTERNAL) &&
1088 percpu_ref_is_dying(&c->writes)) {
1094 * The copygc thread is now global, which means it's no longer
1095 * freeing up space on specific disks, which means that
1096 * allocations for specific disks may hang arbitrarily long:
1098 wp = bch2_alloc_sectors_start(c,
1100 op->opts.erasure_code && !(op->flags & BCH_WRITE_CACHED),
1104 op->nr_replicas_required,
1107 (op->flags & (BCH_WRITE_ALLOC_NOWAIT|
1108 BCH_WRITE_ONLY_SPECIFIED_DEVS)) ? NULL : cl);
1111 if (unlikely(IS_ERR(wp))) {
1112 if (unlikely(PTR_ERR(wp) != -EAGAIN)) {
1121 * It's possible for the allocator to fail, put us on the
1122 * freelist waitlist, and then succeed in one of various retry
1123 * paths: if that happens, we need to disable the skip_put
1124 * optimization because otherwise there won't necessarily be a
1125 * barrier before we free the bch_write_op:
1127 if (atomic_read(&cl->remaining) & CLOSURE_WAITING)
1130 bch2_open_bucket_get(c, wp, &op->open_buckets);
1131 ret = bch2_write_extent(op, wp, &bio);
1132 bch2_alloc_sectors_done(c, wp);
1141 * for the skip_put optimization this has to be set
1142 * before we submit the bio:
1144 op->flags |= BCH_WRITE_DONE;
1147 bio->bi_end_io = bch2_write_endio;
1148 bio->bi_private = &op->cl;
1149 bio->bi_opf |= REQ_OP_WRITE;
1152 closure_get(bio->bi_private);
1154 op->flags |= BCH_WRITE_SKIP_CLOSURE_PUT;
1156 key_to_write = (void *) (op->insert_keys.keys_p +
1157 key_to_write_offset);
1159 bch2_submit_wbio_replicas(to_wbio(bio), c, BCH_DATA_user,
1164 continue_at(cl, bch2_write_index, index_update_wq(op));
1166 memalloc_nofs_restore(nofs_flags);
1170 op->flags |= BCH_WRITE_DONE;
1172 continue_at(cl, bch2_write_index, index_update_wq(op));
1176 * If the write can't all be submitted at once, we generally want to
1177 * block synchronously as that signals backpressure to the caller.
1179 * However, if we're running out of a workqueue, we can't block here
1180 * because we'll be blocking other work items from completing:
1182 if (current->flags & PF_WQ_WORKER) {
1183 continue_at(cl, bch2_write_index, index_update_wq(op));
1189 if (!bch2_keylist_empty(&op->insert_keys)) {
1190 __bch2_write_index(op);
1193 op->flags |= BCH_WRITE_DONE;
1194 continue_at_nobarrier(cl, bch2_write_done, NULL);
1202 static void bch2_write_data_inline(struct bch_write_op *op, unsigned data_len)
1204 struct closure *cl = &op->cl;
1205 struct bio *bio = &op->wbio.bio;
1206 struct bvec_iter iter;
1207 struct bkey_i_inline_data *id;
1211 bch2_check_set_feature(op->c, BCH_FEATURE_inline_data);
1213 ret = bch2_keylist_realloc(&op->insert_keys, op->inline_keys,
1214 ARRAY_SIZE(op->inline_keys),
1215 BKEY_U64s + DIV_ROUND_UP(data_len, 8));
1221 sectors = bio_sectors(bio);
1222 op->pos.offset += sectors;
1224 id = bkey_inline_data_init(op->insert_keys.top);
1226 id->k.version = op->version;
1227 id->k.size = sectors;
1229 iter = bio->bi_iter;
1230 iter.bi_size = data_len;
1231 memcpy_from_bio(id->v.data, bio, iter);
1233 while (data_len & 7)
1234 id->v.data[data_len++] = '\0';
1235 set_bkey_val_bytes(&id->k, data_len);
1236 bch2_keylist_push(&op->insert_keys);
1238 op->flags |= BCH_WRITE_WROTE_DATA_INLINE;
1239 op->flags |= BCH_WRITE_DONE;
1241 continue_at_nobarrier(cl, bch2_write_index, NULL);
1244 bch2_write_done(&op->cl);
1248 * bch_write - handle a write to a cache device or flash only volume
1250 * This is the starting point for any data to end up in a cache device; it could
1251 * be from a normal write, or a writeback write, or a write to a flash only
1252 * volume - it's also used by the moving garbage collector to compact data in
1253 * mostly empty buckets.
1255 * It first writes the data to the cache, creating a list of keys to be inserted
1256 * (if the data won't fit in a single open bucket, there will be multiple keys);
1257 * after the data is written it calls bch_journal, and after the keys have been
1258 * added to the next journal write they're inserted into the btree.
1260 * If op->discard is true, instead of inserting the data it invalidates the
1261 * region of the cache represented by op->bio and op->inode.
1263 void bch2_write(struct closure *cl)
1265 struct bch_write_op *op = container_of(cl, struct bch_write_op, cl);
1266 struct bio *bio = &op->wbio.bio;
1267 struct bch_fs *c = op->c;
1270 BUG_ON(!op->nr_replicas);
1271 BUG_ON(!op->write_point.v);
1272 BUG_ON(!bkey_cmp(op->pos, POS_MAX));
1274 op->start_time = local_clock();
1275 bch2_keylist_init(&op->insert_keys, op->inline_keys);
1276 wbio_init(bio)->put_bio = false;
1278 if (bio->bi_iter.bi_size & (c->opts.block_size - 1)) {
1279 bch_err_inum_ratelimited(c, op->pos.inode,
1280 "misaligned write");
1285 if (c->opts.nochanges ||
1286 !percpu_ref_tryget(&c->writes)) {
1291 bch2_increment_clock(c, bio_sectors(bio), WRITE);
1293 data_len = min_t(u64, bio->bi_iter.bi_size,
1294 op->new_i_size - (op->pos.offset << 9));
1296 if (c->opts.inline_data &&
1297 data_len <= min(block_bytes(c) / 2, 1024U)) {
1298 bch2_write_data_inline(op, data_len);
1302 continue_at_nobarrier(cl, __bch2_write, NULL);
1305 bch2_disk_reservation_put(c, &op->res);
1308 EBUG_ON(cl->parent);
1309 closure_debug_destroy(cl);
1316 /* Cache promotion on read */
1320 struct rcu_head rcu;
1323 struct rhash_head hash;
1326 struct migrate_write write;
1327 struct bio_vec bi_inline_vecs[0]; /* must be last */
1330 static const struct rhashtable_params bch_promote_params = {
1331 .head_offset = offsetof(struct promote_op, hash),
1332 .key_offset = offsetof(struct promote_op, pos),
1333 .key_len = sizeof(struct bpos),
1336 static inline bool should_promote(struct bch_fs *c, struct bkey_s_c k,
1338 struct bch_io_opts opts,
1341 if (!(flags & BCH_READ_MAY_PROMOTE))
1344 if (!opts.promote_target)
1347 if (bch2_bkey_has_target(c, k, opts.promote_target))
1350 if (bch2_target_congested(c, opts.promote_target)) {
1351 /* XXX trace this */
1355 if (rhashtable_lookup_fast(&c->promote_table, &pos,
1356 bch_promote_params))
1362 static void promote_free(struct bch_fs *c, struct promote_op *op)
1366 ret = rhashtable_remove_fast(&c->promote_table, &op->hash,
1367 bch_promote_params);
1369 percpu_ref_put(&c->writes);
1373 static void promote_done(struct closure *cl)
1375 struct promote_op *op =
1376 container_of(cl, struct promote_op, cl);
1377 struct bch_fs *c = op->write.op.c;
1379 bch2_time_stats_update(&c->times[BCH_TIME_data_promote],
1382 bch2_bio_free_pages_pool(c, &op->write.op.wbio.bio);
1383 promote_free(c, op);
1386 static void promote_start(struct promote_op *op, struct bch_read_bio *rbio)
1388 struct bch_fs *c = rbio->c;
1389 struct closure *cl = &op->cl;
1390 struct bio *bio = &op->write.op.wbio.bio;
1392 trace_promote(&rbio->bio);
1394 /* we now own pages: */
1395 BUG_ON(!rbio->bounce);
1396 BUG_ON(rbio->bio.bi_vcnt > bio->bi_max_vecs);
1398 memcpy(bio->bi_io_vec, rbio->bio.bi_io_vec,
1399 sizeof(struct bio_vec) * rbio->bio.bi_vcnt);
1400 swap(bio->bi_vcnt, rbio->bio.bi_vcnt);
1402 bch2_migrate_read_done(&op->write, rbio);
1404 closure_init(cl, NULL);
1405 closure_call(&op->write.op.cl, bch2_write, c->btree_update_wq, cl);
1406 closure_return_with_destructor(cl, promote_done);
1409 static struct promote_op *__promote_alloc(struct bch_fs *c,
1410 enum btree_id btree_id,
1413 struct extent_ptr_decoded *pick,
1414 struct bch_io_opts opts,
1416 struct bch_read_bio **rbio)
1418 struct promote_op *op = NULL;
1420 unsigned pages = DIV_ROUND_UP(sectors, PAGE_SECTORS);
1423 if (!percpu_ref_tryget(&c->writes))
1426 op = kzalloc(sizeof(*op) + sizeof(struct bio_vec) * pages, GFP_NOIO);
1430 op->start_time = local_clock();
1434 * We don't use the mempool here because extents that aren't
1435 * checksummed or compressed can be too big for the mempool:
1437 *rbio = kzalloc(sizeof(struct bch_read_bio) +
1438 sizeof(struct bio_vec) * pages,
1443 rbio_init(&(*rbio)->bio, opts);
1444 bio_init(&(*rbio)->bio, (*rbio)->bio.bi_inline_vecs, pages);
1446 if (bch2_bio_alloc_pages(&(*rbio)->bio, sectors << 9,
1450 (*rbio)->bounce = true;
1451 (*rbio)->split = true;
1452 (*rbio)->kmalloc = true;
1454 if (rhashtable_lookup_insert_fast(&c->promote_table, &op->hash,
1455 bch_promote_params))
1458 bio = &op->write.op.wbio.bio;
1459 bio_init(bio, bio->bi_inline_vecs, pages);
1461 ret = bch2_migrate_write_init(c, &op->write,
1462 writepoint_hashed((unsigned long) current),
1465 (struct data_opts) {
1466 .target = opts.promote_target,
1475 bio_free_pages(&(*rbio)->bio);
1479 percpu_ref_put(&c->writes);
1484 static struct promote_op *promote_alloc(struct bch_fs *c,
1485 struct bvec_iter iter,
1487 struct extent_ptr_decoded *pick,
1488 struct bch_io_opts opts,
1490 struct bch_read_bio **rbio,
1494 bool promote_full = *read_full || READ_ONCE(c->promote_whole_extents);
1495 /* data might have to be decompressed in the write path: */
1496 unsigned sectors = promote_full
1497 ? max(pick->crc.compressed_size, pick->crc.live_size)
1498 : bvec_iter_sectors(iter);
1499 struct bpos pos = promote_full
1500 ? bkey_start_pos(k.k)
1501 : POS(k.k->p.inode, iter.bi_sector);
1502 struct promote_op *promote;
1504 if (!should_promote(c, k, pos, opts, flags))
1507 promote = __promote_alloc(c,
1508 k.k->type == KEY_TYPE_reflink_v
1511 k, pos, pick, opts, sectors, rbio);
1516 *read_full = promote_full;
1522 #define READ_RETRY_AVOID 1
1523 #define READ_RETRY 2
1528 RBIO_CONTEXT_HIGHPRI,
1529 RBIO_CONTEXT_UNBOUND,
1532 static inline struct bch_read_bio *
1533 bch2_rbio_parent(struct bch_read_bio *rbio)
1535 return rbio->split ? rbio->parent : rbio;
1539 static void bch2_rbio_punt(struct bch_read_bio *rbio, work_func_t fn,
1540 enum rbio_context context,
1541 struct workqueue_struct *wq)
1543 if (context <= rbio->context) {
1546 rbio->work.func = fn;
1547 rbio->context = context;
1548 queue_work(wq, &rbio->work);
1552 static inline struct bch_read_bio *bch2_rbio_free(struct bch_read_bio *rbio)
1554 BUG_ON(rbio->bounce && !rbio->split);
1557 promote_free(rbio->c, rbio->promote);
1558 rbio->promote = NULL;
1561 bch2_bio_free_pages_pool(rbio->c, &rbio->bio);
1564 struct bch_read_bio *parent = rbio->parent;
1569 bio_put(&rbio->bio);
1578 * Only called on a top level bch_read_bio to complete an entire read request,
1581 static void bch2_rbio_done(struct bch_read_bio *rbio)
1583 if (rbio->start_time)
1584 bch2_time_stats_update(&rbio->c->times[BCH_TIME_data_read],
1586 bio_endio(&rbio->bio);
1589 static void bch2_read_retry_nodecode(struct bch_fs *c, struct bch_read_bio *rbio,
1590 struct bvec_iter bvec_iter,
1591 struct bch_io_failures *failed,
1594 struct btree_trans trans;
1595 struct btree_iter iter;
1600 flags &= ~BCH_READ_LAST_FRAGMENT;
1601 flags |= BCH_READ_MUST_CLONE;
1603 bch2_bkey_buf_init(&sk);
1604 bch2_trans_init(&trans, c, 0, 0);
1606 bch2_trans_iter_init(&trans, &iter, rbio->data_btree,
1607 rbio->read_pos, BTREE_ITER_SLOTS);
1609 rbio->bio.bi_status = 0;
1611 k = bch2_btree_iter_peek_slot(&iter);
1615 bch2_bkey_buf_reassemble(&sk, c, k);
1616 k = bkey_i_to_s_c(sk.k);
1617 bch2_trans_unlock(&trans);
1619 if (!bch2_bkey_matches_ptr(c, k,
1621 rbio->data_pos.offset -
1622 rbio->pick.crc.offset)) {
1623 /* extent we wanted to read no longer exists: */
1628 ret = __bch2_read_extent(&trans, rbio, bvec_iter,
1631 k, 0, failed, flags);
1632 if (ret == READ_RETRY)
1637 bch2_rbio_done(rbio);
1638 bch2_trans_iter_exit(&trans, &iter);
1639 bch2_trans_exit(&trans);
1640 bch2_bkey_buf_exit(&sk, c);
1643 rbio->bio.bi_status = BLK_STS_IOERR;
1647 static void bch2_rbio_retry(struct work_struct *work)
1649 struct bch_read_bio *rbio =
1650 container_of(work, struct bch_read_bio, work);
1651 struct bch_fs *c = rbio->c;
1652 struct bvec_iter iter = rbio->bvec_iter;
1653 unsigned flags = rbio->flags;
1654 subvol_inum inum = {
1655 .subvol = rbio->subvol,
1656 .inum = rbio->read_pos.inode,
1658 struct bch_io_failures failed = { .nr = 0 };
1660 trace_read_retry(&rbio->bio);
1662 if (rbio->retry == READ_RETRY_AVOID)
1663 bch2_mark_io_failure(&failed, &rbio->pick);
1665 rbio->bio.bi_status = 0;
1667 rbio = bch2_rbio_free(rbio);
1669 flags |= BCH_READ_IN_RETRY;
1670 flags &= ~BCH_READ_MAY_PROMOTE;
1672 if (flags & BCH_READ_NODECODE) {
1673 bch2_read_retry_nodecode(c, rbio, iter, &failed, flags);
1675 flags &= ~BCH_READ_LAST_FRAGMENT;
1676 flags |= BCH_READ_MUST_CLONE;
1678 __bch2_read(c, rbio, iter, inum, &failed, flags);
1682 static void bch2_rbio_error(struct bch_read_bio *rbio, int retry,
1685 rbio->retry = retry;
1687 if (rbio->flags & BCH_READ_IN_RETRY)
1690 if (retry == READ_ERR) {
1691 rbio = bch2_rbio_free(rbio);
1693 rbio->bio.bi_status = error;
1694 bch2_rbio_done(rbio);
1696 bch2_rbio_punt(rbio, bch2_rbio_retry,
1697 RBIO_CONTEXT_UNBOUND, system_unbound_wq);
1701 static int __bch2_rbio_narrow_crcs(struct btree_trans *trans,
1702 struct bch_read_bio *rbio)
1704 struct bch_fs *c = rbio->c;
1705 u64 data_offset = rbio->data_pos.offset - rbio->pick.crc.offset;
1706 struct bch_extent_crc_unpacked new_crc;
1707 struct btree_iter iter;
1712 if (crc_is_compressed(rbio->pick.crc))
1715 bch2_trans_iter_init(trans, &iter, rbio->data_btree, rbio->data_pos,
1716 BTREE_ITER_SLOTS|BTREE_ITER_INTENT);
1717 k = bch2_btree_iter_peek_slot(&iter);
1718 if ((ret = bkey_err(k)))
1721 if (bversion_cmp(k.k->version, rbio->version) ||
1722 !bch2_bkey_matches_ptr(c, k, rbio->pick.ptr, data_offset))
1725 /* Extent was merged? */
1726 if (bkey_start_offset(k.k) < data_offset ||
1727 k.k->p.offset > data_offset + rbio->pick.crc.uncompressed_size)
1730 if (bch2_rechecksum_bio(c, &rbio->bio, rbio->version,
1731 rbio->pick.crc, NULL, &new_crc,
1732 bkey_start_offset(k.k) - data_offset, k.k->size,
1733 rbio->pick.crc.csum_type)) {
1734 bch_err(c, "error verifying existing checksum while narrowing checksum (memory corruption?)");
1740 * going to be temporarily appending another checksum entry:
1742 new = bch2_trans_kmalloc(trans, bkey_bytes(k.k) +
1743 sizeof(struct bch_extent_crc128));
1744 if ((ret = PTR_ERR_OR_ZERO(new)))
1747 bkey_reassemble(new, k);
1749 if (!bch2_bkey_narrow_crcs(new, new_crc))
1752 ret = bch2_trans_update(trans, &iter, new,
1753 BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE);
1755 bch2_trans_iter_exit(trans, &iter);
1759 static noinline void bch2_rbio_narrow_crcs(struct bch_read_bio *rbio)
1761 bch2_trans_do(rbio->c, NULL, NULL, BTREE_INSERT_NOFAIL,
1762 __bch2_rbio_narrow_crcs(&trans, rbio));
1765 /* Inner part that may run in process context */
1766 static void __bch2_read_endio(struct work_struct *work)
1768 struct bch_read_bio *rbio =
1769 container_of(work, struct bch_read_bio, work);
1770 struct bch_fs *c = rbio->c;
1771 struct bch_dev *ca = bch_dev_bkey_exists(c, rbio->pick.ptr.dev);
1772 struct bio *src = &rbio->bio;
1773 struct bio *dst = &bch2_rbio_parent(rbio)->bio;
1774 struct bvec_iter dst_iter = rbio->bvec_iter;
1775 struct bch_extent_crc_unpacked crc = rbio->pick.crc;
1776 struct nonce nonce = extent_nonce(rbio->version, crc);
1777 unsigned nofs_flags;
1778 struct bch_csum csum;
1781 nofs_flags = memalloc_nofs_save();
1783 /* Reset iterator for checksumming and copying bounced data: */
1785 src->bi_iter.bi_size = crc.compressed_size << 9;
1786 src->bi_iter.bi_idx = 0;
1787 src->bi_iter.bi_bvec_done = 0;
1789 src->bi_iter = rbio->bvec_iter;
1792 csum = bch2_checksum_bio(c, crc.csum_type, nonce, src);
1793 if (bch2_crc_cmp(csum, rbio->pick.crc.csum))
1798 * We need to rework the narrow_crcs path to deliver the read completion
1799 * first, and then punt to a different workqueue, otherwise we're
1800 * holding up reads while doing btree updates which is bad for memory
1803 if (unlikely(rbio->narrow_crcs))
1804 bch2_rbio_narrow_crcs(rbio);
1806 if (rbio->flags & BCH_READ_NODECODE)
1809 /* Adjust crc to point to subset of data we want: */
1810 crc.offset += rbio->offset_into_extent;
1811 crc.live_size = bvec_iter_sectors(rbio->bvec_iter);
1813 if (crc_is_compressed(crc)) {
1814 ret = bch2_encrypt_bio(c, crc.csum_type, nonce, src);
1818 if (bch2_bio_uncompress(c, src, dst, dst_iter, crc))
1819 goto decompression_err;
1821 /* don't need to decrypt the entire bio: */
1822 nonce = nonce_add(nonce, crc.offset << 9);
1823 bio_advance(src, crc.offset << 9);
1825 BUG_ON(src->bi_iter.bi_size < dst_iter.bi_size);
1826 src->bi_iter.bi_size = dst_iter.bi_size;
1828 ret = bch2_encrypt_bio(c, crc.csum_type, nonce, src);
1833 struct bvec_iter src_iter = src->bi_iter;
1834 bio_copy_data_iter(dst, &dst_iter, src, &src_iter);
1838 if (rbio->promote) {
1840 * Re encrypt data we decrypted, so it's consistent with
1843 ret = bch2_encrypt_bio(c, crc.csum_type, nonce, src);
1847 promote_start(rbio->promote, rbio);
1848 rbio->promote = NULL;
1851 if (likely(!(rbio->flags & BCH_READ_IN_RETRY))) {
1852 rbio = bch2_rbio_free(rbio);
1853 bch2_rbio_done(rbio);
1856 memalloc_nofs_restore(nofs_flags);
1860 * Checksum error: if the bio wasn't bounced, we may have been
1861 * reading into buffers owned by userspace (that userspace can
1862 * scribble over) - retry the read, bouncing it this time:
1864 if (!rbio->bounce && (rbio->flags & BCH_READ_USER_MAPPED)) {
1865 rbio->flags |= BCH_READ_MUST_BOUNCE;
1866 bch2_rbio_error(rbio, READ_RETRY, BLK_STS_IOERR);
1870 bch2_dev_inum_io_error(ca, rbio->read_pos.inode, (u64) rbio->bvec_iter.bi_sector,
1871 "data checksum error: expected %0llx:%0llx got %0llx:%0llx (type %u)",
1872 rbio->pick.crc.csum.hi, rbio->pick.crc.csum.lo,
1873 csum.hi, csum.lo, crc.csum_type);
1874 bch2_rbio_error(rbio, READ_RETRY_AVOID, BLK_STS_IOERR);
1877 bch_err_inum_ratelimited(c, rbio->read_pos.inode,
1878 "decompression error");
1879 bch2_rbio_error(rbio, READ_ERR, BLK_STS_IOERR);
1882 bch_err_inum_ratelimited(c, rbio->read_pos.inode,
1884 bch2_rbio_error(rbio, READ_ERR, BLK_STS_IOERR);
1888 static void bch2_read_endio(struct bio *bio)
1890 struct bch_read_bio *rbio =
1891 container_of(bio, struct bch_read_bio, bio);
1892 struct bch_fs *c = rbio->c;
1893 struct bch_dev *ca = bch_dev_bkey_exists(c, rbio->pick.ptr.dev);
1894 struct workqueue_struct *wq = NULL;
1895 enum rbio_context context = RBIO_CONTEXT_NULL;
1897 if (rbio->have_ioref) {
1898 bch2_latency_acct(ca, rbio->submit_time, READ);
1899 percpu_ref_put(&ca->io_ref);
1903 rbio->bio.bi_end_io = rbio->end_io;
1905 if (bch2_dev_inum_io_err_on(bio->bi_status, ca,
1906 rbio->read_pos.inode,
1907 rbio->read_pos.offset,
1908 "data read error: %s",
1909 bch2_blk_status_to_str(bio->bi_status))) {
1910 bch2_rbio_error(rbio, READ_RETRY_AVOID, bio->bi_status);
1914 if (((rbio->flags & BCH_READ_RETRY_IF_STALE) && race_fault()) ||
1915 ptr_stale(ca, &rbio->pick.ptr)) {
1916 atomic_long_inc(&c->read_realloc_races);
1918 if (rbio->flags & BCH_READ_RETRY_IF_STALE)
1919 bch2_rbio_error(rbio, READ_RETRY, BLK_STS_AGAIN);
1921 bch2_rbio_error(rbio, READ_ERR, BLK_STS_AGAIN);
1925 if (rbio->narrow_crcs ||
1926 crc_is_compressed(rbio->pick.crc) ||
1927 bch2_csum_type_is_encryption(rbio->pick.crc.csum_type))
1928 context = RBIO_CONTEXT_UNBOUND, wq = system_unbound_wq;
1929 else if (rbio->pick.crc.csum_type)
1930 context = RBIO_CONTEXT_HIGHPRI, wq = system_highpri_wq;
1932 bch2_rbio_punt(rbio, __bch2_read_endio, context, wq);
1935 int __bch2_read_indirect_extent(struct btree_trans *trans,
1936 unsigned *offset_into_extent,
1937 struct bkey_buf *orig_k)
1939 struct btree_iter iter;
1944 reflink_offset = le64_to_cpu(bkey_i_to_reflink_p(orig_k->k)->v.idx) +
1945 *offset_into_extent;
1947 bch2_trans_iter_init(trans, &iter, BTREE_ID_reflink,
1948 POS(0, reflink_offset),
1950 k = bch2_btree_iter_peek_slot(&iter);
1955 if (k.k->type != KEY_TYPE_reflink_v &&
1956 k.k->type != KEY_TYPE_indirect_inline_data) {
1957 bch_err_inum_ratelimited(trans->c, orig_k->k->k.p.inode,
1958 "%llu len %u points to nonexistent indirect extent %llu",
1959 orig_k->k->k.p.offset,
1962 bch2_inconsistent_error(trans->c);
1967 *offset_into_extent = iter.pos.offset - bkey_start_offset(k.k);
1968 bch2_bkey_buf_reassemble(orig_k, trans->c, k);
1970 bch2_trans_iter_exit(trans, &iter);
1974 static noinline void read_from_stale_dirty_pointer(struct btree_trans *trans,
1976 struct bch_extent_ptr ptr)
1978 struct bch_fs *c = trans->c;
1979 struct bch_dev *ca = bch_dev_bkey_exists(c, ptr.dev);
1980 struct btree_iter iter;
1981 struct printbuf buf = PRINTBUF;
1984 bch2_trans_iter_init(trans, &iter, BTREE_ID_alloc,
1985 PTR_BUCKET_POS(c, &ptr),
1988 pr_buf(&buf, "Attempting to read from stale dirty pointer:");
1989 pr_indent_push(&buf, 2);
1992 bch2_bkey_val_to_text(&buf, c, k);
1995 pr_buf(&buf, "memory gen: %u", *bucket_gen(ca, iter.pos.offset));
1997 ret = lockrestart_do(trans, bkey_err(k = bch2_btree_iter_peek_slot(&iter)));
2000 bch2_bkey_val_to_text(&buf, c, k);
2003 bch2_fs_inconsistent(c, "%s", buf.buf);
2005 bch2_trans_iter_exit(trans, &iter);
2006 printbuf_exit(&buf);
2009 int __bch2_read_extent(struct btree_trans *trans, struct bch_read_bio *orig,
2010 struct bvec_iter iter, struct bpos read_pos,
2011 enum btree_id data_btree, struct bkey_s_c k,
2012 unsigned offset_into_extent,
2013 struct bch_io_failures *failed, unsigned flags)
2015 struct bch_fs *c = trans->c;
2016 struct extent_ptr_decoded pick;
2017 struct bch_read_bio *rbio = NULL;
2018 struct bch_dev *ca = NULL;
2019 struct promote_op *promote = NULL;
2020 bool bounce = false, read_full = false, narrow_crcs = false;
2021 struct bpos data_pos = bkey_start_pos(k.k);
2024 if (bkey_extent_is_inline_data(k.k)) {
2025 unsigned bytes = min_t(unsigned, iter.bi_size,
2026 bkey_inline_data_bytes(k.k));
2028 swap(iter.bi_size, bytes);
2029 memcpy_to_bio(&orig->bio, iter, bkey_inline_data_p(k));
2030 swap(iter.bi_size, bytes);
2031 bio_advance_iter(&orig->bio, &iter, bytes);
2032 zero_fill_bio_iter(&orig->bio, iter);
2036 pick_ret = bch2_bkey_pick_read_device(c, k, failed, &pick);
2038 /* hole or reservation - just zero fill: */
2043 bch_err_inum_ratelimited(c, k.k->p.inode,
2044 "no device to read from");
2048 ca = bch_dev_bkey_exists(c, pick.ptr.dev);
2051 * Stale dirty pointers are treated as IO errors, but @failed isn't
2052 * allocated unless we're in the retry path - so if we're not in the
2053 * retry path, don't check here, it'll be caught in bch2_read_endio()
2054 * and we'll end up in the retry path:
2056 if ((flags & BCH_READ_IN_RETRY) &&
2058 unlikely(ptr_stale(ca, &pick.ptr))) {
2059 read_from_stale_dirty_pointer(trans, k, pick.ptr);
2060 bch2_mark_io_failure(failed, &pick);
2065 * Unlock the iterator while the btree node's lock is still in
2066 * cache, before doing the IO:
2068 bch2_trans_unlock(trans);
2070 if (flags & BCH_READ_NODECODE) {
2072 * can happen if we retry, and the extent we were going to read
2073 * has been merged in the meantime:
2075 if (pick.crc.compressed_size > orig->bio.bi_vcnt * PAGE_SECTORS)
2078 iter.bi_size = pick.crc.compressed_size << 9;
2082 if (!(flags & BCH_READ_LAST_FRAGMENT) ||
2083 bio_flagged(&orig->bio, BIO_CHAIN))
2084 flags |= BCH_READ_MUST_CLONE;
2086 narrow_crcs = !(flags & BCH_READ_IN_RETRY) &&
2087 bch2_can_narrow_extent_crcs(k, pick.crc);
2089 if (narrow_crcs && (flags & BCH_READ_USER_MAPPED))
2090 flags |= BCH_READ_MUST_BOUNCE;
2092 EBUG_ON(offset_into_extent + bvec_iter_sectors(iter) > k.k->size);
2094 if (crc_is_compressed(pick.crc) ||
2095 (pick.crc.csum_type != BCH_CSUM_none &&
2096 (bvec_iter_sectors(iter) != pick.crc.uncompressed_size ||
2097 (bch2_csum_type_is_encryption(pick.crc.csum_type) &&
2098 (flags & BCH_READ_USER_MAPPED)) ||
2099 (flags & BCH_READ_MUST_BOUNCE)))) {
2104 if (orig->opts.promote_target)
2105 promote = promote_alloc(c, iter, k, &pick, orig->opts, flags,
2106 &rbio, &bounce, &read_full);
2109 EBUG_ON(crc_is_compressed(pick.crc));
2110 EBUG_ON(pick.crc.csum_type &&
2111 (bvec_iter_sectors(iter) != pick.crc.uncompressed_size ||
2112 bvec_iter_sectors(iter) != pick.crc.live_size ||
2114 offset_into_extent));
2116 data_pos.offset += offset_into_extent;
2117 pick.ptr.offset += pick.crc.offset +
2119 offset_into_extent = 0;
2120 pick.crc.compressed_size = bvec_iter_sectors(iter);
2121 pick.crc.uncompressed_size = bvec_iter_sectors(iter);
2122 pick.crc.offset = 0;
2123 pick.crc.live_size = bvec_iter_sectors(iter);
2124 offset_into_extent = 0;
2129 * promote already allocated bounce rbio:
2130 * promote needs to allocate a bio big enough for uncompressing
2131 * data in the write path, but we're not going to use it all
2134 EBUG_ON(rbio->bio.bi_iter.bi_size <
2135 pick.crc.compressed_size << 9);
2136 rbio->bio.bi_iter.bi_size =
2137 pick.crc.compressed_size << 9;
2138 } else if (bounce) {
2139 unsigned sectors = pick.crc.compressed_size;
2141 rbio = rbio_init(bio_alloc_bioset(GFP_NOIO,
2142 DIV_ROUND_UP(sectors, PAGE_SECTORS),
2143 &c->bio_read_split),
2146 bch2_bio_alloc_pages_pool(c, &rbio->bio, sectors << 9);
2147 rbio->bounce = true;
2149 } else if (flags & BCH_READ_MUST_CLONE) {
2151 * Have to clone if there were any splits, due to error
2152 * reporting issues (if a split errored, and retrying didn't
2153 * work, when it reports the error to its parent (us) we don't
2154 * know if the error was from our bio, and we should retry, or
2155 * from the whole bio, in which case we don't want to retry and
2158 rbio = rbio_init(bio_clone_fast(&orig->bio, GFP_NOIO,
2159 &c->bio_read_split),
2161 rbio->bio.bi_iter = iter;
2165 rbio->bio.bi_iter = iter;
2166 EBUG_ON(bio_flagged(&rbio->bio, BIO_CHAIN));
2169 EBUG_ON(bio_sectors(&rbio->bio) != pick.crc.compressed_size);
2172 rbio->submit_time = local_clock();
2174 rbio->parent = orig;
2176 rbio->end_io = orig->bio.bi_end_io;
2177 rbio->bvec_iter = iter;
2178 rbio->offset_into_extent= offset_into_extent;
2179 rbio->flags = flags;
2180 rbio->have_ioref = pick_ret > 0 && bch2_dev_get_ioref(ca, READ);
2181 rbio->narrow_crcs = narrow_crcs;
2185 /* XXX: only initialize this if needed */
2186 rbio->devs_have = bch2_bkey_devs(k);
2188 rbio->subvol = orig->subvol;
2189 rbio->read_pos = read_pos;
2190 rbio->data_btree = data_btree;
2191 rbio->data_pos = data_pos;
2192 rbio->version = k.k->version;
2193 rbio->promote = promote;
2194 INIT_WORK(&rbio->work, NULL);
2196 rbio->bio.bi_opf = orig->bio.bi_opf;
2197 rbio->bio.bi_iter.bi_sector = pick.ptr.offset;
2198 rbio->bio.bi_end_io = bch2_read_endio;
2201 trace_read_bounce(&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 if (!bch2_trans_relock(&trans)) {
2329 bch2_btree_iter_set_pos(&iter,
2330 POS(inum.inum, bvec_iter.bi_sector));
2332 k = bch2_btree_iter_peek_slot(&iter);
2337 offset_into_extent = iter.pos.offset -
2338 bkey_start_offset(k.k);
2339 sectors = k.k->size - offset_into_extent;
2341 bch2_bkey_buf_reassemble(&sk, c, k);
2343 ret = bch2_read_indirect_extent(&trans, &data_btree,
2344 &offset_into_extent, &sk);
2348 k = bkey_i_to_s_c(sk.k);
2351 * With indirect extents, the amount of data to read is the min
2352 * of the original extent and the indirect extent:
2354 sectors = min(sectors, k.k->size - offset_into_extent);
2356 bytes = min(sectors, bvec_iter_sectors(bvec_iter)) << 9;
2357 swap(bvec_iter.bi_size, bytes);
2359 if (bvec_iter.bi_size == bytes)
2360 flags |= BCH_READ_LAST_FRAGMENT;
2362 ret = __bch2_read_extent(&trans, rbio, bvec_iter, iter.pos,
2364 offset_into_extent, failed, flags);
2368 if (flags & BCH_READ_LAST_FRAGMENT)
2371 swap(bvec_iter.bi_size, bytes);
2372 bio_advance_iter(&rbio->bio, &bvec_iter, bytes);
2374 ret = btree_trans_too_many_iters(&trans);
2379 bch2_trans_iter_exit(&trans, &iter);
2381 if (ret == -EINTR || ret == READ_RETRY || ret == READ_RETRY_AVOID)
2384 bch2_trans_exit(&trans);
2385 bch2_bkey_buf_exit(&sk, c);
2388 bch_err_inum_ratelimited(c, inum.inum,
2389 "read error %i from btree lookup", ret);
2390 rbio->bio.bi_status = BLK_STS_IOERR;
2391 bch2_rbio_done(rbio);
2395 void bch2_fs_io_exit(struct bch_fs *c)
2397 if (c->promote_table.tbl)
2398 rhashtable_destroy(&c->promote_table);
2399 mempool_exit(&c->bio_bounce_pages);
2400 bioset_exit(&c->bio_write);
2401 bioset_exit(&c->bio_read_split);
2402 bioset_exit(&c->bio_read);
2405 int bch2_fs_io_init(struct bch_fs *c)
2407 if (bioset_init(&c->bio_read, 1, offsetof(struct bch_read_bio, bio),
2408 BIOSET_NEED_BVECS) ||
2409 bioset_init(&c->bio_read_split, 1, offsetof(struct bch_read_bio, bio),
2410 BIOSET_NEED_BVECS) ||
2411 bioset_init(&c->bio_write, 1, offsetof(struct bch_write_bio, bio),
2412 BIOSET_NEED_BVECS) ||
2413 mempool_init_page_pool(&c->bio_bounce_pages,
2415 c->opts.btree_node_size,
2416 c->opts.encoded_extent_max) /
2418 rhashtable_init(&c->promote_table, &bch_promote_params))