2 * Some low level IO code, and hacks for various block layer limitations
4 * Copyright 2010, 2011 Kent Overstreet <kent.overstreet@gmail.com>
5 * Copyright 2012 Google, Inc.
11 #include "btree_update.h"
17 #include "disk_groups.h"
24 #include "rebalance.h"
29 #include <linux/blkdev.h>
30 #include <linux/random.h>
32 #include <trace/events/bcachefs.h>
34 static bool bch2_target_congested(struct bch_fs *c, u16 target)
36 const struct bch_devs_mask *devs;
37 unsigned d, nr = 0, total = 0;
38 u64 now = local_clock(), last;
46 devs = bch2_target_to_mask(c, target);
47 for_each_set_bit(d, devs->d, BCH_SB_MEMBERS_MAX) {
48 ca = rcu_dereference(c->devs[d]);
52 congested = atomic_read(&ca->congested);
53 last = READ_ONCE(ca->congested_last);
54 if (time_after64(now, last))
55 congested -= (now - last) >> 12;
57 total += max(congested, 0LL);
62 return bch2_rand_range(nr * CONGESTED_MAX) < total;
65 static inline void bch2_congested_acct(struct bch_dev *ca, u64 io_latency,
69 ca->io_latency[rw].quantiles.entries[QUANTILE_IDX(1)].m;
70 /* ideally we'd be taking into account the device's variance here: */
71 u64 latency_threshold = latency_capable << (rw == READ ? 2 : 3);
72 s64 latency_over = io_latency - latency_threshold;
74 if (latency_threshold && latency_over > 0) {
76 * bump up congested by approximately latency_over * 4 /
77 * latency_threshold - we don't need much accuracy here so don't
78 * bother with the divide:
80 if (atomic_read(&ca->congested) < CONGESTED_MAX)
81 atomic_add(latency_over >>
82 max_t(int, ilog2(latency_threshold) - 2, 0),
85 ca->congested_last = now;
86 } else if (atomic_read(&ca->congested) > 0) {
87 atomic_dec(&ca->congested);
91 void bch2_latency_acct(struct bch_dev *ca, u64 submit_time, int rw)
93 atomic64_t *latency = &ca->cur_latency[rw];
94 u64 now = local_clock();
95 u64 io_latency = time_after64(now, submit_time)
98 u64 old, new, v = atomic64_read(latency);
104 * If the io latency was reasonably close to the current
105 * latency, skip doing the update and atomic operation - most of
108 if (abs((int) (old - io_latency)) < (old >> 1) &&
112 new = ewma_add(old, io_latency, 5);
113 } while ((v = atomic64_cmpxchg(latency, old, new)) != old);
115 bch2_congested_acct(ca, io_latency, now, rw);
117 __bch2_time_stats_update(&ca->io_latency[rw], submit_time, now);
120 /* Allocate, free from mempool: */
122 void bch2_bio_free_pages_pool(struct bch_fs *c, struct bio *bio)
127 bio_for_each_segment_all(bv, bio, i)
128 if (bv->bv_page != ZERO_PAGE(0))
129 mempool_free(bv->bv_page, &c->bio_bounce_pages);
133 static void bch2_bio_alloc_page_pool(struct bch_fs *c, struct bio *bio,
136 struct bio_vec *bv = &bio->bi_io_vec[bio->bi_vcnt++];
138 if (likely(!*using_mempool)) {
139 bv->bv_page = alloc_page(GFP_NOIO);
140 if (unlikely(!bv->bv_page)) {
141 mutex_lock(&c->bio_bounce_pages_lock);
142 *using_mempool = true;
148 bv->bv_page = mempool_alloc(&c->bio_bounce_pages, GFP_NOIO);
151 bv->bv_len = PAGE_SIZE;
155 void bch2_bio_alloc_pages_pool(struct bch_fs *c, struct bio *bio,
158 bool using_mempool = false;
160 BUG_ON(DIV_ROUND_UP(bytes, PAGE_SIZE) > bio->bi_max_vecs);
162 bio->bi_iter.bi_size = bytes;
164 while (bio->bi_vcnt < DIV_ROUND_UP(bytes, PAGE_SIZE))
165 bch2_bio_alloc_page_pool(c, bio, &using_mempool);
168 mutex_unlock(&c->bio_bounce_pages_lock);
171 void bch2_bio_alloc_more_pages_pool(struct bch_fs *c, struct bio *bio,
174 while (bio->bi_vcnt < DIV_ROUND_UP(bytes, PAGE_SIZE)) {
175 struct bio_vec *bv = &bio->bi_io_vec[bio->bi_vcnt];
177 BUG_ON(bio->bi_vcnt >= bio->bi_max_vecs);
179 bv->bv_page = alloc_page(GFP_NOIO);
182 * We already allocated from mempool, we can't allocate from it again
183 * without freeing the pages we already allocated or else we could
186 bch2_bio_free_pages_pool(c, bio);
187 bch2_bio_alloc_pages_pool(c, bio, bytes);
191 bv->bv_len = PAGE_SIZE;
196 bio->bi_iter.bi_size = bytes;
201 void bch2_submit_wbio_replicas(struct bch_write_bio *wbio, struct bch_fs *c,
202 enum bch_data_type type,
203 const struct bkey_i *k)
205 struct bkey_s_c_extent e = bkey_i_to_s_c_extent(k);
206 const struct bch_extent_ptr *ptr;
207 struct bch_write_bio *n;
210 BUG_ON(c->opts.nochanges);
212 extent_for_each_ptr(e, ptr) {
213 BUG_ON(ptr->dev >= BCH_SB_MEMBERS_MAX ||
216 ca = bch_dev_bkey_exists(c, ptr->dev);
218 if (ptr + 1 < &extent_entry_last(e)->ptr) {
219 n = to_wbio(bio_clone_fast(&wbio->bio, GFP_NOIO,
222 n->bio.bi_end_io = wbio->bio.bi_end_io;
223 n->bio.bi_private = wbio->bio.bi_private;
228 n->bio.bi_opf = wbio->bio.bi_opf;
229 bio_inc_remaining(&wbio->bio);
237 n->have_ioref = bch2_dev_get_ioref(ca, WRITE);
238 n->submit_time = local_clock();
239 n->bio.bi_iter.bi_sector = ptr->offset;
241 if (!journal_flushes_device(ca))
242 n->bio.bi_opf |= REQ_FUA;
244 if (likely(n->have_ioref)) {
245 this_cpu_add(ca->io_done->sectors[WRITE][type],
246 bio_sectors(&n->bio));
248 bio_set_dev(&n->bio, ca->disk_sb.bdev);
251 n->bio.bi_status = BLK_STS_REMOVED;
257 static void __bch2_write(struct closure *);
259 static void bch2_write_done(struct closure *cl)
261 struct bch_write_op *op = container_of(cl, struct bch_write_op, cl);
262 struct bch_fs *c = op->c;
264 if (!op->error && (op->flags & BCH_WRITE_FLUSH))
265 op->error = bch2_journal_error(&c->journal);
267 if (!(op->flags & BCH_WRITE_NOPUT_RESERVATION))
268 bch2_disk_reservation_put(c, &op->res);
269 percpu_ref_put(&c->writes);
270 bch2_keylist_free(&op->insert_keys, op->inline_keys);
272 bch2_time_stats_update(&c->times[BCH_TIME_data_write], op->start_time);
277 int bch2_write_index_default(struct bch_write_op *op)
279 struct keylist *keys = &op->insert_keys;
280 struct btree_iter iter;
283 bch2_btree_iter_init(&iter, op->c, BTREE_ID_EXTENTS,
284 bkey_start_pos(&bch2_keylist_front(keys)->k),
287 ret = bch2_btree_insert_list_at(&iter, keys, &op->res,
290 BTREE_INSERT_USE_RESERVE);
291 bch2_btree_iter_unlock(&iter);
297 * bch_write_index - after a write, update index to point to new data
299 static void __bch2_write_index(struct bch_write_op *op)
301 struct bch_fs *c = op->c;
302 struct keylist *keys = &op->insert_keys;
303 struct bkey_s_extent e;
304 struct bch_extent_ptr *ptr;
305 struct bkey_i *src, *dst = keys->keys, *n, *k;
308 for (src = keys->keys; src != keys->top; src = n) {
312 e = bkey_i_to_s_extent(dst);
313 extent_for_each_ptr_backwards(e, ptr)
314 if (test_bit(ptr->dev, op->failed.d))
315 bch2_extent_drop_ptr(e, ptr);
317 if (!bch2_extent_nr_ptrs(e.c)) {
322 if (!(op->flags & BCH_WRITE_NOMARK_REPLICAS)) {
323 ret = bch2_mark_bkey_replicas(c, BCH_DATA_USER, e.s_c);
328 dst = bkey_next(dst);
334 * probably not the ideal place to hook this in, but I don't
335 * particularly want to plumb io_opts all the way through the btree
336 * update stack right now
338 for_each_keylist_key(keys, k)
339 bch2_rebalance_add_key(c, bkey_i_to_s_c(k), &op->opts);
341 if (!bch2_keylist_empty(keys)) {
342 u64 sectors_start = keylist_sectors(keys);
343 int ret = op->index_update_fn(op);
345 BUG_ON(keylist_sectors(keys) && !ret);
347 op->written += sectors_start - keylist_sectors(keys);
350 __bcache_io_error(c, "btree IO error %i", ret);
355 bch2_open_bucket_put_refs(c, &op->open_buckets_nr, op->open_buckets);
358 keys->top = keys->keys;
363 static void bch2_write_index(struct closure *cl)
365 struct bch_write_op *op = container_of(cl, struct bch_write_op, cl);
366 struct bch_fs *c = op->c;
368 __bch2_write_index(op);
370 if (!op->error && (op->flags & BCH_WRITE_FLUSH)) {
371 bch2_journal_flush_seq_async(&c->journal,
374 continue_at(cl, bch2_write_done, index_update_wq(op));
376 continue_at_nobarrier(cl, bch2_write_done, NULL);
380 static void bch2_write_endio(struct bio *bio)
382 struct closure *cl = bio->bi_private;
383 struct bch_write_op *op = container_of(cl, struct bch_write_op, cl);
384 struct bch_write_bio *wbio = to_wbio(bio);
385 struct bch_write_bio *parent = wbio->split ? wbio->parent : NULL;
386 struct bch_fs *c = wbio->c;
387 struct bch_dev *ca = bch_dev_bkey_exists(c, wbio->dev);
389 if (bch2_dev_io_err_on(bio->bi_status, ca, "data write"))
390 set_bit(wbio->dev, op->failed.d);
392 if (wbio->have_ioref) {
393 bch2_latency_acct(ca, wbio->submit_time, WRITE);
394 percpu_ref_put(&ca->io_ref);
398 bch2_bio_free_pages_pool(c, bio);
404 bio_endio(&parent->bio);
409 static void init_append_extent(struct bch_write_op *op,
410 struct write_point *wp,
411 struct bversion version,
412 struct bch_extent_crc_unpacked crc)
414 struct bkey_i_extent *e = bkey_extent_init(op->insert_keys.top);
416 op->pos.offset += crc.uncompressed_size;
418 e->k.size = crc.uncompressed_size;
419 e->k.version = version;
420 bkey_extent_set_cached(&e->k, op->flags & BCH_WRITE_CACHED);
422 bch2_extent_crc_append(e, crc);
423 bch2_alloc_sectors_append_ptrs(op->c, wp, e, crc.compressed_size);
425 bch2_keylist_push(&op->insert_keys);
428 static struct bio *bch2_write_bio_alloc(struct bch_fs *c,
429 struct write_point *wp,
431 bool *page_alloc_failed)
433 struct bch_write_bio *wbio;
435 unsigned output_available =
436 min(wp->sectors_free << 9, src->bi_iter.bi_size);
437 unsigned pages = DIV_ROUND_UP(output_available, PAGE_SIZE);
439 bio = bio_alloc_bioset(GFP_NOIO, pages, &c->bio_write);
440 wbio = wbio_init(bio);
442 wbio->put_bio = true;
443 /* copy WRITE_SYNC flag */
444 wbio->bio.bi_opf = src->bi_opf;
447 * We can't use mempool for more than c->sb.encoded_extent_max
448 * worth of pages, but we'd like to allocate more if we can:
450 while (bio->bi_iter.bi_size < output_available) {
451 unsigned len = min_t(unsigned, PAGE_SIZE,
452 output_available - bio->bi_iter.bi_size);
455 p = alloc_page(GFP_NOIO);
458 min_t(unsigned, output_available,
459 c->sb.encoded_extent_max << 9);
461 if (bio_sectors(bio) < pool_max)
462 bch2_bio_alloc_pages_pool(c, bio, pool_max);
466 bio->bi_io_vec[bio->bi_vcnt++] = (struct bio_vec) {
471 bio->bi_iter.bi_size += len;
474 *page_alloc_failed = bio->bi_vcnt < pages;
478 static int bch2_write_rechecksum(struct bch_fs *c,
479 struct bch_write_op *op,
480 unsigned new_csum_type)
482 struct bio *bio = &op->wbio.bio;
483 struct bch_extent_crc_unpacked new_crc;
486 /* bch2_rechecksum_bio() can't encrypt or decrypt data: */
488 if (bch2_csum_type_is_encryption(op->crc.csum_type) !=
489 bch2_csum_type_is_encryption(new_csum_type))
490 new_csum_type = op->crc.csum_type;
492 ret = bch2_rechecksum_bio(c, bio, op->version, op->crc,
494 op->crc.offset, op->crc.live_size,
499 bio_advance(bio, op->crc.offset << 9);
500 bio->bi_iter.bi_size = op->crc.live_size << 9;
505 static int bch2_write_decrypt(struct bch_write_op *op)
507 struct bch_fs *c = op->c;
508 struct nonce nonce = extent_nonce(op->version, op->crc);
509 struct bch_csum csum;
511 if (!bch2_csum_type_is_encryption(op->crc.csum_type))
515 * If we need to decrypt data in the write path, we'll no longer be able
516 * to verify the existing checksum (poly1305 mac, in this case) after
517 * it's decrypted - this is the last point we'll be able to reverify the
520 csum = bch2_checksum_bio(c, op->crc.csum_type, nonce, &op->wbio.bio);
521 if (bch2_crc_cmp(op->crc.csum, csum))
524 bch2_encrypt_bio(c, op->crc.csum_type, nonce, &op->wbio.bio);
525 op->crc.csum_type = 0;
526 op->crc.csum = (struct bch_csum) { 0, 0 };
530 static enum prep_encoded_ret {
533 PREP_ENCODED_CHECKSUM_ERR,
534 PREP_ENCODED_DO_WRITE,
535 } bch2_write_prep_encoded_data(struct bch_write_op *op, struct write_point *wp)
537 struct bch_fs *c = op->c;
538 struct bio *bio = &op->wbio.bio;
540 if (!(op->flags & BCH_WRITE_DATA_ENCODED))
541 return PREP_ENCODED_OK;
543 BUG_ON(bio_sectors(bio) != op->crc.compressed_size);
545 /* Can we just write the entire extent as is? */
546 if (op->crc.uncompressed_size == op->crc.live_size &&
547 op->crc.compressed_size <= wp->sectors_free &&
548 op->crc.compression_type == op->compression_type) {
549 if (!op->crc.compression_type &&
550 op->csum_type != op->crc.csum_type &&
551 bch2_write_rechecksum(c, op, op->csum_type))
552 return PREP_ENCODED_CHECKSUM_ERR;
554 return PREP_ENCODED_DO_WRITE;
558 * If the data is compressed and we couldn't write the entire extent as
559 * is, we have to decompress it:
561 if (op->crc.compression_type) {
562 struct bch_csum csum;
564 if (bch2_write_decrypt(op))
565 return PREP_ENCODED_CHECKSUM_ERR;
567 /* Last point we can still verify checksum: */
568 csum = bch2_checksum_bio(c, op->crc.csum_type,
569 extent_nonce(op->version, op->crc),
571 if (bch2_crc_cmp(op->crc.csum, csum))
572 return PREP_ENCODED_CHECKSUM_ERR;
574 if (bch2_bio_uncompress_inplace(c, bio, &op->crc))
575 return PREP_ENCODED_ERR;
579 * No longer have compressed data after this point - data might be
584 * If the data is checksummed and we're only writing a subset,
585 * rechecksum and adjust bio to point to currently live data:
587 if ((op->crc.live_size != op->crc.uncompressed_size ||
588 op->crc.csum_type != op->csum_type) &&
589 bch2_write_rechecksum(c, op, op->csum_type))
590 return PREP_ENCODED_CHECKSUM_ERR;
593 * If we want to compress the data, it has to be decrypted:
595 if ((op->compression_type ||
596 bch2_csum_type_is_encryption(op->crc.csum_type) !=
597 bch2_csum_type_is_encryption(op->csum_type)) &&
598 bch2_write_decrypt(op))
599 return PREP_ENCODED_CHECKSUM_ERR;
601 return PREP_ENCODED_OK;
604 static int bch2_write_extent(struct bch_write_op *op, struct write_point *wp)
606 struct bch_fs *c = op->c;
607 struct bio *src = &op->wbio.bio, *dst = src;
608 struct bvec_iter saved_iter;
609 struct bkey_i *key_to_write;
610 unsigned key_to_write_offset = op->insert_keys.top_p -
611 op->insert_keys.keys_p;
612 unsigned total_output = 0;
613 bool bounce = false, page_alloc_failed = false;
616 BUG_ON(!bio_sectors(src));
618 switch (bch2_write_prep_encoded_data(op, wp)) {
619 case PREP_ENCODED_OK:
621 case PREP_ENCODED_ERR:
624 case PREP_ENCODED_CHECKSUM_ERR:
626 case PREP_ENCODED_DO_WRITE:
627 init_append_extent(op, wp, op->version, op->crc);
631 if (op->compression_type ||
633 !(op->flags & BCH_WRITE_PAGES_STABLE)) ||
634 (bch2_csum_type_is_encryption(op->csum_type) &&
635 !(op->flags & BCH_WRITE_PAGES_OWNED))) {
636 dst = bch2_write_bio_alloc(c, wp, src, &page_alloc_failed);
640 saved_iter = dst->bi_iter;
643 struct bch_extent_crc_unpacked crc =
644 (struct bch_extent_crc_unpacked) { 0 };
645 struct bversion version = op->version;
646 size_t dst_len, src_len;
648 if (page_alloc_failed &&
649 bio_sectors(dst) < wp->sectors_free &&
650 bio_sectors(dst) < c->sb.encoded_extent_max)
653 BUG_ON(op->compression_type &&
654 (op->flags & BCH_WRITE_DATA_ENCODED) &&
655 bch2_csum_type_is_encryption(op->crc.csum_type));
656 BUG_ON(op->compression_type && !bounce);
658 crc.compression_type = op->compression_type
659 ? bch2_bio_compress(c, dst, &dst_len, src, &src_len,
660 op->compression_type)
662 if (!crc.compression_type) {
663 dst_len = min(dst->bi_iter.bi_size, src->bi_iter.bi_size);
664 dst_len = min_t(unsigned, dst_len, wp->sectors_free << 9);
667 dst_len = min_t(unsigned, dst_len,
668 c->sb.encoded_extent_max << 9);
671 swap(dst->bi_iter.bi_size, dst_len);
672 bio_copy_data(dst, src);
673 swap(dst->bi_iter.bi_size, dst_len);
679 BUG_ON(!src_len || !dst_len);
681 if (bch2_csum_type_is_encryption(op->csum_type)) {
682 if (bversion_zero(version)) {
683 version.lo = atomic64_inc_return(&c->key_version) + 1;
685 crc.nonce = op->nonce;
686 op->nonce += src_len >> 9;
690 if ((op->flags & BCH_WRITE_DATA_ENCODED) &&
691 !crc.compression_type &&
692 bch2_csum_type_is_encryption(op->crc.csum_type) ==
693 bch2_csum_type_is_encryption(op->csum_type)) {
695 * Note: when we're using rechecksum(), we need to be
696 * checksumming @src because it has all the data our
697 * existing checksum covers - if we bounced (because we
698 * were trying to compress), @dst will only have the
699 * part of the data the new checksum will cover.
701 * But normally we want to be checksumming post bounce,
702 * because part of the reason for bouncing is so the
703 * data can't be modified (by userspace) while it's in
706 if (bch2_rechecksum_bio(c, src, version, op->crc,
709 bio_sectors(src) - (src_len >> 9),
713 if ((op->flags & BCH_WRITE_DATA_ENCODED) &&
714 bch2_rechecksum_bio(c, src, version, op->crc,
717 bio_sectors(src) - (src_len >> 9),
721 crc.compressed_size = dst_len >> 9;
722 crc.uncompressed_size = src_len >> 9;
723 crc.live_size = src_len >> 9;
725 swap(dst->bi_iter.bi_size, dst_len);
726 bch2_encrypt_bio(c, op->csum_type,
727 extent_nonce(version, crc), dst);
728 crc.csum = bch2_checksum_bio(c, op->csum_type,
729 extent_nonce(version, crc), dst);
730 crc.csum_type = op->csum_type;
731 swap(dst->bi_iter.bi_size, dst_len);
734 init_append_extent(op, wp, version, crc);
737 bio_advance(dst, dst_len);
738 bio_advance(src, src_len);
739 total_output += dst_len;
740 } while (dst->bi_iter.bi_size &&
741 src->bi_iter.bi_size &&
743 !bch2_keylist_realloc(&op->insert_keys,
745 ARRAY_SIZE(op->inline_keys),
746 BKEY_EXTENT_U64s_MAX));
748 more = src->bi_iter.bi_size != 0;
750 dst->bi_iter = saved_iter;
752 if (!bounce && more) {
753 dst = bio_split(src, total_output >> 9,
754 GFP_NOIO, &c->bio_write);
755 wbio_init(dst)->put_bio = true;
758 dst->bi_iter.bi_size = total_output;
760 /* Free unneeded pages after compressing: */
762 while (dst->bi_vcnt > DIV_ROUND_UP(dst->bi_iter.bi_size, PAGE_SIZE))
763 mempool_free(dst->bi_io_vec[--dst->bi_vcnt].bv_page,
764 &c->bio_bounce_pages);
766 /* might have done a realloc... */
768 key_to_write = (void *) (op->insert_keys.keys_p + key_to_write_offset);
770 dst->bi_end_io = bch2_write_endio;
771 dst->bi_private = &op->cl;
772 bio_set_op_attrs(dst, REQ_OP_WRITE, 0);
774 closure_get(dst->bi_private);
776 bch2_submit_wbio_replicas(to_wbio(dst), c, BCH_DATA_USER,
780 bch_err(c, "error verifying existing checksum while "
781 "rewriting existing data (memory corruption?)");
785 bch2_bio_free_pages_pool(c, dst);
792 static void __bch2_write(struct closure *cl)
794 struct bch_write_op *op = container_of(cl, struct bch_write_op, cl);
795 struct bch_fs *c = op->c;
796 struct write_point *wp;
800 /* +1 for possible cache device: */
801 if (op->open_buckets_nr + op->nr_replicas + 1 >
802 ARRAY_SIZE(op->open_buckets))
805 if (bch2_keylist_realloc(&op->insert_keys,
807 ARRAY_SIZE(op->inline_keys),
808 BKEY_EXTENT_U64s_MAX))
811 wp = bch2_alloc_sectors_start(c,
816 op->nr_replicas_required,
819 (op->flags & BCH_WRITE_ALLOC_NOWAIT) ? NULL : cl);
822 if (unlikely(IS_ERR(wp))) {
823 if (unlikely(PTR_ERR(wp) != -EAGAIN)) {
831 ret = bch2_write_extent(op, wp);
833 BUG_ON(op->open_buckets_nr + wp->nr_ptrs - wp->first_ptr >
834 ARRAY_SIZE(op->open_buckets));
835 bch2_open_bucket_get(c, wp,
836 &op->open_buckets_nr,
838 bch2_alloc_sectors_done(c, wp);
844 continue_at(cl, bch2_write_index, index_update_wq(op));
849 continue_at(cl, !bch2_keylist_empty(&op->insert_keys)
851 : bch2_write_done, index_update_wq(op));
856 if (!bch2_keylist_empty(&op->insert_keys)) {
857 __bch2_write_index(op);
860 continue_at_nobarrier(cl, bch2_write_done, NULL);
869 * bch_write - handle a write to a cache device or flash only volume
871 * This is the starting point for any data to end up in a cache device; it could
872 * be from a normal write, or a writeback write, or a write to a flash only
873 * volume - it's also used by the moving garbage collector to compact data in
874 * mostly empty buckets.
876 * It first writes the data to the cache, creating a list of keys to be inserted
877 * (if the data won't fit in a single open bucket, there will be multiple keys);
878 * after the data is written it calls bch_journal, and after the keys have been
879 * added to the next journal write they're inserted into the btree.
881 * If op->discard is true, instead of inserting the data it invalidates the
882 * region of the cache represented by op->bio and op->inode.
884 void bch2_write(struct closure *cl)
886 struct bch_write_op *op = container_of(cl, struct bch_write_op, cl);
887 struct bch_fs *c = op->c;
889 BUG_ON(!op->nr_replicas);
890 BUG_ON(!op->write_point.v);
891 BUG_ON(!bkey_cmp(op->pos, POS_MAX));
892 BUG_ON(bio_sectors(&op->wbio.bio) > U16_MAX);
894 op->start_time = local_clock();
896 memset(&op->failed, 0, sizeof(op->failed));
898 bch2_keylist_init(&op->insert_keys, op->inline_keys);
899 wbio_init(&op->wbio.bio)->put_bio = false;
901 if (c->opts.nochanges ||
902 !percpu_ref_tryget(&c->writes)) {
903 __bcache_io_error(c, "read only");
905 if (!(op->flags & BCH_WRITE_NOPUT_RESERVATION))
906 bch2_disk_reservation_put(c, &op->res);
911 bch2_increment_clock(c, bio_sectors(&op->wbio.bio), WRITE);
913 continue_at_nobarrier(cl, __bch2_write, NULL);
916 /* Cache promotion on read */
922 struct rhash_head hash;
925 struct migrate_write write;
926 struct bio_vec bi_inline_vecs[0]; /* must be last */
929 static const struct rhashtable_params bch_promote_params = {
930 .head_offset = offsetof(struct promote_op, hash),
931 .key_offset = offsetof(struct promote_op, pos),
932 .key_len = sizeof(struct bpos),
935 static inline bool should_promote(struct bch_fs *c, struct bkey_s_c k,
937 struct bch_io_opts opts,
940 if (!opts.promote_target)
943 if (!(flags & BCH_READ_MAY_PROMOTE))
946 if (percpu_ref_is_dying(&c->writes))
949 if (!bkey_extent_is_data(k.k))
952 if (bch2_extent_has_target(c, bkey_s_c_to_extent(k), opts.promote_target))
955 if (bch2_target_congested(c, opts.promote_target))
958 if (rhashtable_lookup_fast(&c->promote_table, &pos,
965 static void promote_free(struct bch_fs *c, struct promote_op *op)
969 ret = rhashtable_remove_fast(&c->promote_table, &op->hash,
972 percpu_ref_put(&c->writes);
976 static void promote_done(struct closure *cl)
978 struct promote_op *op =
979 container_of(cl, struct promote_op, cl);
980 struct bch_fs *c = op->write.op.c;
982 bch2_time_stats_update(&c->times[BCH_TIME_data_promote],
985 bch2_bio_free_pages_pool(c, &op->write.op.wbio.bio);
989 static void promote_start(struct promote_op *op, struct bch_read_bio *rbio)
991 struct bch_fs *c = rbio->c;
992 struct closure *cl = &op->cl;
993 struct bio *bio = &op->write.op.wbio.bio;
995 trace_promote(&rbio->bio);
997 /* we now own pages: */
998 BUG_ON(!rbio->bounce);
999 BUG_ON(rbio->bio.bi_vcnt > bio->bi_max_vecs);
1001 memcpy(bio->bi_io_vec, rbio->bio.bi_io_vec,
1002 sizeof(struct bio_vec) * rbio->bio.bi_vcnt);
1003 swap(bio->bi_vcnt, rbio->bio.bi_vcnt);
1005 bch2_migrate_read_done(&op->write, rbio);
1007 closure_init(cl, NULL);
1008 closure_call(&op->write.op.cl, bch2_write, c->wq, cl);
1009 closure_return_with_destructor(cl, promote_done);
1013 static struct promote_op *__promote_alloc(struct bch_fs *c,
1015 struct extent_pick_ptr *pick,
1016 struct bch_io_opts opts,
1017 unsigned rbio_sectors,
1018 struct bch_read_bio **rbio)
1020 struct promote_op *op = NULL;
1022 unsigned rbio_pages = DIV_ROUND_UP(rbio_sectors, PAGE_SECTORS);
1023 /* data might have to be decompressed in the write path: */
1024 unsigned wbio_pages = DIV_ROUND_UP(pick->crc.uncompressed_size,
1028 if (!percpu_ref_tryget(&c->writes))
1031 op = kzalloc(sizeof(*op) + sizeof(struct bio_vec) * wbio_pages,
1036 op->start_time = local_clock();
1040 * promotes require bouncing, but if the extent isn't
1041 * checksummed/compressed it might be too big for the mempool:
1043 if (rbio_sectors > c->sb.encoded_extent_max) {
1044 *rbio = kzalloc(sizeof(struct bch_read_bio) +
1045 sizeof(struct bio_vec) * rbio_pages,
1050 rbio_init(&(*rbio)->bio, opts);
1051 bio_init(&(*rbio)->bio, (*rbio)->bio.bi_inline_vecs,
1054 (*rbio)->bio.bi_iter.bi_size = rbio_sectors << 9;
1055 bch2_bio_map(&(*rbio)->bio, NULL);
1057 if (bch2_bio_alloc_pages(&(*rbio)->bio, GFP_NOIO))
1060 (*rbio)->bounce = true;
1061 (*rbio)->split = true;
1062 (*rbio)->kmalloc = true;
1065 if (rhashtable_lookup_insert_fast(&c->promote_table, &op->hash,
1066 bch_promote_params))
1069 bio = &op->write.op.wbio.bio;
1070 bio_init(bio, bio->bi_inline_vecs, wbio_pages);
1072 ret = bch2_migrate_write_init(c, &op->write,
1073 writepoint_hashed((unsigned long) current),
1076 (struct data_opts) {
1077 .target = opts.promote_target
1085 bio_free_pages(&(*rbio)->bio);
1089 percpu_ref_put(&c->writes);
1093 static inline struct promote_op *promote_alloc(struct bch_fs *c,
1094 struct bvec_iter iter,
1096 struct extent_pick_ptr *pick,
1097 struct bch_io_opts opts,
1099 struct bch_read_bio **rbio,
1103 bool promote_full = *read_full || READ_ONCE(c->promote_whole_extents);
1104 unsigned sectors = promote_full
1105 ? pick->crc.compressed_size
1106 : bvec_iter_sectors(iter);
1107 struct bpos pos = promote_full
1108 ? bkey_start_pos(k.k)
1109 : POS(k.k->p.inode, iter.bi_sector);
1110 struct promote_op *promote;
1112 if (!should_promote(c, k, pos, opts, flags))
1115 promote = __promote_alloc(c, pos, pick, opts, sectors, rbio);
1120 *read_full = promote_full;
1126 #define READ_RETRY_AVOID 1
1127 #define READ_RETRY 2
1132 RBIO_CONTEXT_HIGHPRI,
1133 RBIO_CONTEXT_UNBOUND,
1136 static inline struct bch_read_bio *
1137 bch2_rbio_parent(struct bch_read_bio *rbio)
1139 return rbio->split ? rbio->parent : rbio;
1143 static void bch2_rbio_punt(struct bch_read_bio *rbio, work_func_t fn,
1144 enum rbio_context context,
1145 struct workqueue_struct *wq)
1147 if (context <= rbio->context) {
1150 rbio->work.func = fn;
1151 rbio->context = context;
1152 queue_work(wq, &rbio->work);
1156 static inline struct bch_read_bio *bch2_rbio_free(struct bch_read_bio *rbio)
1158 BUG_ON(rbio->bounce && !rbio->split);
1161 promote_free(rbio->c, rbio->promote);
1162 rbio->promote = NULL;
1165 bch2_bio_free_pages_pool(rbio->c, &rbio->bio);
1168 struct bch_read_bio *parent = rbio->parent;
1173 bio_put(&rbio->bio);
1181 static void bch2_rbio_done(struct bch_read_bio *rbio)
1183 bch2_time_stats_update(&rbio->c->times[BCH_TIME_data_read],
1185 bio_endio(&rbio->bio);
1188 static void bch2_read_retry_nodecode(struct bch_fs *c, struct bch_read_bio *rbio,
1189 struct bvec_iter bvec_iter, u64 inode,
1190 struct bch_devs_mask *avoid, unsigned flags)
1192 struct btree_iter iter;
1197 flags &= ~BCH_READ_LAST_FRAGMENT;
1199 bch2_btree_iter_init(&iter, c, BTREE_ID_EXTENTS,
1200 rbio->pos, BTREE_ITER_SLOTS);
1202 rbio->bio.bi_status = 0;
1204 k = bch2_btree_iter_peek_slot(&iter);
1205 if (btree_iter_err(k)) {
1206 bch2_btree_iter_unlock(&iter);
1210 bkey_reassemble(&tmp.k, k);
1211 k = bkey_i_to_s_c(&tmp.k);
1212 bch2_btree_iter_unlock(&iter);
1214 if (!bkey_extent_is_data(k.k) ||
1215 !bch2_extent_matches_ptr(c, bkey_i_to_s_c_extent(&tmp.k),
1218 rbio->pick.crc.offset)) {
1219 /* extent we wanted to read no longer exists: */
1224 ret = __bch2_read_extent(c, rbio, bvec_iter, k, avoid, flags);
1225 if (ret == READ_RETRY)
1231 rbio->bio.bi_status = BLK_STS_IOERR;
1233 bch2_rbio_done(rbio);
1236 static void bch2_read_retry(struct bch_fs *c, struct bch_read_bio *rbio,
1237 struct bvec_iter bvec_iter, u64 inode,
1238 struct bch_devs_mask *avoid, unsigned flags)
1240 struct btree_iter iter;
1244 flags &= ~BCH_READ_LAST_FRAGMENT;
1245 flags |= BCH_READ_MUST_CLONE;
1247 for_each_btree_key(&iter, c, BTREE_ID_EXTENTS,
1248 POS(inode, bvec_iter.bi_sector),
1249 BTREE_ITER_SLOTS, k) {
1253 bkey_reassemble(&tmp.k, k);
1254 k = bkey_i_to_s_c(&tmp.k);
1255 bch2_btree_iter_unlock(&iter);
1257 bytes = min_t(unsigned, bvec_iter.bi_size,
1258 (k.k->p.offset - bvec_iter.bi_sector) << 9);
1259 swap(bvec_iter.bi_size, bytes);
1261 ret = __bch2_read_extent(c, rbio, bvec_iter, k, avoid, flags);
1269 if (bytes == bvec_iter.bi_size)
1272 swap(bvec_iter.bi_size, bytes);
1273 bio_advance_iter(&rbio->bio, &bvec_iter, bytes);
1277 * If we get here, it better have been because there was an error
1278 * reading a btree node
1280 ret = bch2_btree_iter_unlock(&iter);
1282 __bcache_io_error(c, "btree IO error %i", ret);
1284 rbio->bio.bi_status = BLK_STS_IOERR;
1286 bch2_rbio_done(rbio);
1289 static void bch2_rbio_retry(struct work_struct *work)
1291 struct bch_read_bio *rbio =
1292 container_of(work, struct bch_read_bio, work);
1293 struct bch_fs *c = rbio->c;
1294 struct bvec_iter iter = rbio->bvec_iter;
1295 unsigned flags = rbio->flags;
1296 u64 inode = rbio->pos.inode;
1297 struct bch_devs_mask avoid;
1299 trace_read_retry(&rbio->bio);
1301 memset(&avoid, 0, sizeof(avoid));
1303 if (rbio->retry == READ_RETRY_AVOID)
1304 __set_bit(rbio->pick.ptr.dev, avoid.d);
1306 rbio->bio.bi_status = 0;
1308 rbio = bch2_rbio_free(rbio);
1310 flags |= BCH_READ_IN_RETRY;
1311 flags &= ~BCH_READ_MAY_PROMOTE;
1313 if (flags & BCH_READ_NODECODE)
1314 bch2_read_retry_nodecode(c, rbio, iter, inode, &avoid, flags);
1316 bch2_read_retry(c, rbio, iter, inode, &avoid, flags);
1319 static void bch2_rbio_error(struct bch_read_bio *rbio, int retry,
1322 rbio->retry = retry;
1324 if (rbio->flags & BCH_READ_IN_RETRY)
1327 if (retry == READ_ERR) {
1328 rbio = bch2_rbio_free(rbio);
1330 rbio->bio.bi_status = error;
1331 bch2_rbio_done(rbio);
1333 bch2_rbio_punt(rbio, bch2_rbio_retry,
1334 RBIO_CONTEXT_UNBOUND, system_unbound_wq);
1338 static void bch2_rbio_narrow_crcs(struct bch_read_bio *rbio)
1340 struct bch_fs *c = rbio->c;
1341 struct btree_iter iter;
1343 struct bkey_i_extent *e;
1345 struct bch_extent_crc_unpacked new_crc;
1349 if (rbio->pick.crc.compression_type)
1352 bch2_btree_iter_init(&iter, c, BTREE_ID_EXTENTS, rbio->pos,
1355 k = bch2_btree_iter_peek(&iter);
1356 if (IS_ERR_OR_NULL(k.k))
1359 if (!bkey_extent_is_data(k.k))
1362 bkey_reassemble(&new.k, k);
1363 e = bkey_i_to_extent(&new.k);
1365 if (!bch2_extent_matches_ptr(c, extent_i_to_s_c(e),
1368 rbio->pick.crc.offset) ||
1369 bversion_cmp(e->k.version, rbio->version))
1372 /* Extent was merged? */
1373 if (bkey_start_offset(&e->k) < rbio->pos.offset ||
1374 e->k.p.offset > rbio->pos.offset + rbio->pick.crc.uncompressed_size)
1377 /* The extent might have been partially overwritten since we read it: */
1378 offset = rbio->pick.crc.offset + (bkey_start_offset(&e->k) - rbio->pos.offset);
1380 if (bch2_rechecksum_bio(c, &rbio->bio, rbio->version,
1381 rbio->pick.crc, NULL, &new_crc,
1383 rbio->pick.crc.csum_type)) {
1384 bch_err(c, "error verifying existing checksum while narrowing checksum (memory corruption?)");
1388 if (!bch2_extent_narrow_crcs(e, new_crc))
1391 ret = bch2_btree_insert_at(c, NULL, NULL,
1392 BTREE_INSERT_ATOMIC|
1393 BTREE_INSERT_NOFAIL|
1394 BTREE_INSERT_NOWAIT,
1395 BTREE_INSERT_ENTRY(&iter, &e->k_i));
1399 bch2_btree_iter_unlock(&iter);
1402 static bool should_narrow_crcs(struct bkey_s_c k,
1403 struct extent_pick_ptr *pick,
1406 return !(flags & BCH_READ_IN_RETRY) &&
1407 bkey_extent_is_data(k.k) &&
1408 bch2_can_narrow_extent_crcs(bkey_s_c_to_extent(k), pick->crc);
1411 /* Inner part that may run in process context */
1412 static void __bch2_read_endio(struct work_struct *work)
1414 struct bch_read_bio *rbio =
1415 container_of(work, struct bch_read_bio, work);
1416 struct bch_fs *c = rbio->c;
1417 struct bch_dev *ca = bch_dev_bkey_exists(c, rbio->pick.ptr.dev);
1418 struct bio *src = &rbio->bio;
1419 struct bio *dst = &bch2_rbio_parent(rbio)->bio;
1420 struct bvec_iter dst_iter = rbio->bvec_iter;
1421 struct bch_extent_crc_unpacked crc = rbio->pick.crc;
1422 struct nonce nonce = extent_nonce(rbio->version, crc);
1423 struct bch_csum csum;
1425 /* Reset iterator for checksumming and copying bounced data: */
1427 src->bi_iter.bi_size = crc.compressed_size << 9;
1428 src->bi_iter.bi_idx = 0;
1429 src->bi_iter.bi_bvec_done = 0;
1431 src->bi_iter = rbio->bvec_iter;
1434 csum = bch2_checksum_bio(c, crc.csum_type, nonce, src);
1435 if (bch2_crc_cmp(csum, rbio->pick.crc.csum))
1438 if (unlikely(rbio->narrow_crcs))
1439 bch2_rbio_narrow_crcs(rbio);
1441 if (rbio->flags & BCH_READ_NODECODE)
1444 /* Adjust crc to point to subset of data we want: */
1445 crc.offset += rbio->bvec_iter.bi_sector - rbio->pos.offset;
1446 crc.live_size = bvec_iter_sectors(rbio->bvec_iter);
1448 if (crc.compression_type != BCH_COMPRESSION_NONE) {
1449 bch2_encrypt_bio(c, crc.csum_type, nonce, src);
1450 if (bch2_bio_uncompress(c, src, dst, dst_iter, crc))
1451 goto decompression_err;
1453 /* don't need to decrypt the entire bio: */
1454 nonce = nonce_add(nonce, crc.offset << 9);
1455 bio_advance(src, crc.offset << 9);
1457 BUG_ON(src->bi_iter.bi_size < dst_iter.bi_size);
1458 src->bi_iter.bi_size = dst_iter.bi_size;
1460 bch2_encrypt_bio(c, crc.csum_type, nonce, src);
1463 struct bvec_iter src_iter = src->bi_iter;
1464 bio_copy_data_iter(dst, &dst_iter, src, &src_iter);
1468 if (rbio->promote) {
1470 * Re encrypt data we decrypted, so it's consistent with
1473 bch2_encrypt_bio(c, crc.csum_type, nonce, src);
1474 promote_start(rbio->promote, rbio);
1475 rbio->promote = NULL;
1478 if (likely(!(rbio->flags & BCH_READ_IN_RETRY))) {
1479 rbio = bch2_rbio_free(rbio);
1480 bch2_rbio_done(rbio);
1485 * Checksum error: if the bio wasn't bounced, we may have been
1486 * reading into buffers owned by userspace (that userspace can
1487 * scribble over) - retry the read, bouncing it this time:
1489 if (!rbio->bounce && (rbio->flags & BCH_READ_USER_MAPPED)) {
1490 rbio->flags |= BCH_READ_MUST_BOUNCE;
1491 bch2_rbio_error(rbio, READ_RETRY, BLK_STS_IOERR);
1495 bch2_dev_io_error(ca,
1496 "data checksum error, inode %llu offset %llu: expected %0llx:%0llx got %0llx:%0llx (type %u)",
1497 rbio->pos.inode, (u64) rbio->bvec_iter.bi_sector,
1498 rbio->pick.crc.csum.hi, rbio->pick.crc.csum.lo,
1499 csum.hi, csum.lo, crc.csum_type);
1500 bch2_rbio_error(rbio, READ_RETRY_AVOID, BLK_STS_IOERR);
1503 __bcache_io_error(c, "decompression error, inode %llu offset %llu",
1505 (u64) rbio->bvec_iter.bi_sector);
1506 bch2_rbio_error(rbio, READ_ERR, BLK_STS_IOERR);
1510 static void bch2_read_endio(struct bio *bio)
1512 struct bch_read_bio *rbio =
1513 container_of(bio, struct bch_read_bio, bio);
1514 struct bch_fs *c = rbio->c;
1515 struct bch_dev *ca = bch_dev_bkey_exists(c, rbio->pick.ptr.dev);
1516 struct workqueue_struct *wq = NULL;
1517 enum rbio_context context = RBIO_CONTEXT_NULL;
1519 if (rbio->have_ioref) {
1520 bch2_latency_acct(ca, rbio->submit_time, READ);
1521 percpu_ref_put(&ca->io_ref);
1525 rbio->bio.bi_end_io = rbio->end_io;
1527 if (bch2_dev_io_err_on(bio->bi_status, ca, "data read")) {
1528 bch2_rbio_error(rbio, READ_RETRY_AVOID, bio->bi_status);
1532 if (rbio->pick.ptr.cached &&
1533 (((rbio->flags & BCH_READ_RETRY_IF_STALE) && race_fault()) ||
1534 ptr_stale(ca, &rbio->pick.ptr))) {
1535 atomic_long_inc(&c->read_realloc_races);
1537 if (rbio->flags & BCH_READ_RETRY_IF_STALE)
1538 bch2_rbio_error(rbio, READ_RETRY, BLK_STS_AGAIN);
1540 bch2_rbio_error(rbio, READ_ERR, BLK_STS_AGAIN);
1544 if (rbio->narrow_crcs ||
1545 rbio->pick.crc.compression_type ||
1546 bch2_csum_type_is_encryption(rbio->pick.crc.csum_type))
1547 context = RBIO_CONTEXT_UNBOUND, wq = system_unbound_wq;
1548 else if (rbio->pick.crc.csum_type)
1549 context = RBIO_CONTEXT_HIGHPRI, wq = system_highpri_wq;
1551 bch2_rbio_punt(rbio, __bch2_read_endio, context, wq);
1554 int __bch2_read_extent(struct bch_fs *c, struct bch_read_bio *orig,
1555 struct bvec_iter iter, struct bkey_s_c k,
1556 struct bch_devs_mask *avoid, unsigned flags)
1558 struct extent_pick_ptr pick;
1559 struct bch_read_bio *rbio = NULL;
1561 struct promote_op *promote = NULL;
1562 bool bounce = false, read_full = false, narrow_crcs = false;
1563 struct bpos pos = bkey_start_pos(k.k);
1566 pick_ret = bch2_extent_pick_ptr(c, k, avoid, &pick);
1568 /* hole or reservation - just zero fill: */
1576 ca = bch_dev_bkey_exists(c, pick.ptr.dev);
1578 if (flags & BCH_READ_NODECODE) {
1580 * can happen if we retry, and the extent we were going to read
1581 * has been merged in the meantime:
1583 if (pick.crc.compressed_size > orig->bio.bi_vcnt * PAGE_SECTORS)
1586 iter.bi_sector = pos.offset;
1587 iter.bi_size = pick.crc.compressed_size << 9;
1591 if (!(flags & BCH_READ_LAST_FRAGMENT) ||
1592 bio_flagged(&orig->bio, BIO_CHAIN))
1593 flags |= BCH_READ_MUST_CLONE;
1595 narrow_crcs = should_narrow_crcs(k, &pick, flags);
1597 if (narrow_crcs && (flags & BCH_READ_USER_MAPPED))
1598 flags |= BCH_READ_MUST_BOUNCE;
1600 EBUG_ON(bkey_start_offset(k.k) > iter.bi_sector ||
1601 k.k->p.offset < bvec_iter_end_sector(iter));
1603 if (pick.crc.compression_type != BCH_COMPRESSION_NONE ||
1604 (pick.crc.csum_type != BCH_CSUM_NONE &&
1605 (bvec_iter_sectors(iter) != pick.crc.uncompressed_size ||
1606 (bch2_csum_type_is_encryption(pick.crc.csum_type) &&
1607 (flags & BCH_READ_USER_MAPPED)) ||
1608 (flags & BCH_READ_MUST_BOUNCE)))) {
1613 promote = promote_alloc(c, iter, k, &pick, orig->opts, flags,
1614 &rbio, &bounce, &read_full);
1617 EBUG_ON(pick.crc.compression_type);
1618 EBUG_ON(pick.crc.csum_type &&
1619 (bvec_iter_sectors(iter) != pick.crc.uncompressed_size ||
1620 bvec_iter_sectors(iter) != pick.crc.live_size ||
1622 iter.bi_sector != pos.offset));
1624 pick.ptr.offset += pick.crc.offset +
1625 (iter.bi_sector - pos.offset);
1626 pick.crc.compressed_size = bvec_iter_sectors(iter);
1627 pick.crc.uncompressed_size = bvec_iter_sectors(iter);
1628 pick.crc.offset = 0;
1629 pick.crc.live_size = bvec_iter_sectors(iter);
1630 pos.offset = iter.bi_sector;
1634 /* promote already allocated bounce rbio */
1635 } else if (bounce) {
1636 unsigned sectors = pick.crc.compressed_size;
1638 rbio = rbio_init(bio_alloc_bioset(GFP_NOIO,
1639 DIV_ROUND_UP(sectors, PAGE_SECTORS),
1640 &c->bio_read_split),
1643 bch2_bio_alloc_pages_pool(c, &rbio->bio, sectors << 9);
1644 rbio->bounce = true;
1646 } else if (flags & BCH_READ_MUST_CLONE) {
1648 * Have to clone if there were any splits, due to error
1649 * reporting issues (if a split errored, and retrying didn't
1650 * work, when it reports the error to its parent (us) we don't
1651 * know if the error was from our bio, and we should retry, or
1652 * from the whole bio, in which case we don't want to retry and
1655 rbio = rbio_init(bio_clone_fast(&orig->bio, GFP_NOIO,
1656 &c->bio_read_split),
1658 rbio->bio.bi_iter = iter;
1663 rbio->bio.bi_iter = iter;
1664 BUG_ON(bio_flagged(&rbio->bio, BIO_CHAIN));
1667 BUG_ON(bio_sectors(&rbio->bio) != pick.crc.compressed_size);
1670 rbio->submit_time = local_clock();
1672 rbio->parent = orig;
1674 rbio->end_io = orig->bio.bi_end_io;
1675 rbio->bvec_iter = iter;
1676 rbio->flags = flags;
1677 rbio->have_ioref = pick_ret > 0 && bch2_dev_get_ioref(ca, READ);
1678 rbio->narrow_crcs = narrow_crcs;
1682 rbio->devs_have = bch2_bkey_devs(k);
1685 rbio->version = k.k->version;
1686 rbio->promote = promote;
1687 INIT_WORK(&rbio->work, NULL);
1689 rbio->bio.bi_opf = orig->bio.bi_opf;
1690 rbio->bio.bi_iter.bi_sector = pick.ptr.offset;
1691 rbio->bio.bi_end_io = bch2_read_endio;
1694 trace_read_bounce(&rbio->bio);
1696 bch2_increment_clock(c, bio_sectors(&rbio->bio), READ);
1698 if (!rbio->have_ioref)
1699 goto no_device_postclone;
1701 percpu_down_read_preempt_disable(&c->usage_lock);
1702 bucket_io_clock_reset(c, ca, PTR_BUCKET_NR(ca, &pick.ptr), READ);
1703 percpu_up_read_preempt_enable(&c->usage_lock);
1705 this_cpu_add(ca->io_done->sectors[READ][BCH_DATA_USER],
1706 bio_sectors(&rbio->bio));
1708 bio_set_dev(&rbio->bio, ca->disk_sb.bdev);
1710 if (likely(!(flags & BCH_READ_IN_RETRY))) {
1711 if (!(flags & BCH_READ_LAST_FRAGMENT)) {
1712 bio_inc_remaining(&orig->bio);
1713 trace_read_split(&orig->bio);
1716 submit_bio(&rbio->bio);
1721 submit_bio_wait(&rbio->bio);
1723 rbio->context = RBIO_CONTEXT_UNBOUND;
1724 bch2_read_endio(&rbio->bio);
1727 rbio = bch2_rbio_free(rbio);
1729 if (ret == READ_RETRY_AVOID) {
1730 __set_bit(pick.ptr.dev, avoid->d);
1737 no_device_postclone:
1739 rbio->bio.bi_end_io = rbio->end_io;
1740 bch2_rbio_free(rbio);
1742 __bcache_io_error(c, "no device to read from");
1744 if (likely(!(flags & BCH_READ_IN_RETRY))) {
1745 orig->bio.bi_status = BLK_STS_IOERR;
1747 if (flags & BCH_READ_LAST_FRAGMENT)
1748 bch2_rbio_done(orig);
1756 * won't normally happen in the BCH_READ_NODECODE
1757 * (bch2_move_extent()) path, but if we retry and the extent we wanted
1758 * to read no longer exists we have to signal that:
1760 if (flags & BCH_READ_NODECODE)
1763 zero_fill_bio_iter(&orig->bio, iter);
1765 if (flags & BCH_READ_LAST_FRAGMENT)
1766 bch2_rbio_done(orig);
1770 void bch2_read(struct bch_fs *c, struct bch_read_bio *rbio, u64 inode)
1772 struct btree_iter iter;
1774 unsigned flags = BCH_READ_RETRY_IF_STALE|
1775 BCH_READ_MAY_PROMOTE|
1776 BCH_READ_USER_MAPPED;
1779 BUG_ON(rbio->_state);
1780 BUG_ON(flags & BCH_READ_NODECODE);
1781 BUG_ON(flags & BCH_READ_IN_RETRY);
1784 rbio->start_time = local_clock();
1786 for_each_btree_key(&iter, c, BTREE_ID_EXTENTS,
1787 POS(inode, rbio->bio.bi_iter.bi_sector),
1788 BTREE_ITER_SLOTS, k) {
1793 * Unlock the iterator while the btree node's lock is still in
1794 * cache, before doing the IO:
1796 bkey_reassemble(&tmp.k, k);
1797 k = bkey_i_to_s_c(&tmp.k);
1798 bch2_btree_iter_unlock(&iter);
1800 bytes = min_t(unsigned, rbio->bio.bi_iter.bi_size,
1801 (k.k->p.offset - rbio->bio.bi_iter.bi_sector) << 9);
1802 swap(rbio->bio.bi_iter.bi_size, bytes);
1804 if (rbio->bio.bi_iter.bi_size == bytes)
1805 flags |= BCH_READ_LAST_FRAGMENT;
1807 bch2_read_extent(c, rbio, k, flags);
1809 if (flags & BCH_READ_LAST_FRAGMENT)
1812 swap(rbio->bio.bi_iter.bi_size, bytes);
1813 bio_advance(&rbio->bio, bytes);
1817 * If we get here, it better have been because there was an error
1818 * reading a btree node
1820 ret = bch2_btree_iter_unlock(&iter);
1822 bcache_io_error(c, &rbio->bio, "btree IO error %i", ret);
1823 bch2_rbio_done(rbio);
1826 void bch2_fs_io_exit(struct bch_fs *c)
1828 if (c->promote_table.tbl)
1829 rhashtable_destroy(&c->promote_table);
1830 mempool_exit(&c->bio_bounce_pages);
1831 bioset_exit(&c->bio_write);
1832 bioset_exit(&c->bio_read_split);
1833 bioset_exit(&c->bio_read);
1836 int bch2_fs_io_init(struct bch_fs *c)
1838 if (bioset_init(&c->bio_read, 1, offsetof(struct bch_read_bio, bio),
1839 BIOSET_NEED_BVECS) ||
1840 bioset_init(&c->bio_read_split, 1, offsetof(struct bch_read_bio, bio),
1841 BIOSET_NEED_BVECS) ||
1842 bioset_init(&c->bio_write, 1, offsetof(struct bch_write_bio, bio),
1843 BIOSET_NEED_BVECS) ||
1844 mempool_init_page_pool(&c->bio_bounce_pages,
1846 c->opts.btree_node_size,
1847 c->sb.encoded_extent_max) /
1849 rhashtable_init(&c->promote_table, &bch_promote_params))