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.
9 #include "alloc_foreground.h"
11 #include "btree_update.h"
17 #include "disk_groups.h"
25 #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_ptrs_c ptrs = bch2_bkey_ptrs_c(bkey_i_to_s_c(k));
206 const struct bch_extent_ptr *ptr;
207 struct bch_write_bio *n;
210 BUG_ON(c->opts.nochanges);
212 bkey_for_each_ptr(ptrs, ptr) {
213 BUG_ON(ptr->dev >= BCH_SB_MEMBERS_MAX ||
216 ca = bch_dev_bkey_exists(c, ptr->dev);
218 if (to_entry(ptr + 1) < ptrs.end) {
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 bch_fs *c = op->c;
280 struct btree_trans trans;
281 struct btree_iter *iter;
282 struct keylist *keys = &op->insert_keys;
285 BUG_ON(bch2_keylist_empty(keys));
286 bch2_verify_keylist_sorted(keys);
288 bch2_trans_init(&trans, c);
290 iter = bch2_trans_get_iter(&trans, BTREE_ID_EXTENTS,
291 bkey_start_pos(&bch2_keylist_front(keys)->k),
295 BKEY_PADDED(k) split;
297 bkey_copy(&split.k, bch2_keylist_front(keys));
299 bch2_extent_trim_atomic(&split.k, iter);
301 bch2_trans_update(&trans,
302 BTREE_INSERT_ENTRY(iter, &split.k));
304 ret = bch2_trans_commit(&trans, &op->res, op_journal_seq(op),
306 BTREE_INSERT_USE_RESERVE);
310 if (bkey_cmp(iter->pos, bch2_keylist_front(keys)->k.p) < 0)
311 bch2_cut_front(iter->pos, bch2_keylist_front(keys));
313 bch2_keylist_pop_front(keys);
314 } while (!bch2_keylist_empty(keys));
316 bch2_trans_exit(&trans);
322 * bch_write_index - after a write, update index to point to new data
324 static void __bch2_write_index(struct bch_write_op *op)
326 struct bch_fs *c = op->c;
327 struct keylist *keys = &op->insert_keys;
328 struct bch_extent_ptr *ptr;
329 struct bkey_i *src, *dst = keys->keys, *n, *k;
333 for (src = keys->keys; src != keys->top; src = n) {
337 bch2_bkey_drop_ptrs(bkey_i_to_s(dst), ptr,
338 test_bit(ptr->dev, op->failed.d));
340 if (!bch2_bkey_nr_ptrs(bkey_i_to_s_c(dst))) {
345 dst = bkey_next(dst);
351 * probably not the ideal place to hook this in, but I don't
352 * particularly want to plumb io_opts all the way through the btree
353 * update stack right now
355 for_each_keylist_key(keys, k)
356 bch2_rebalance_add_key(c, bkey_i_to_s_c(k), &op->opts);
358 if (!bch2_keylist_empty(keys)) {
359 u64 sectors_start = keylist_sectors(keys);
360 int ret = op->index_update_fn(op);
362 BUG_ON(keylist_sectors(keys) && !ret);
364 op->written += sectors_start - keylist_sectors(keys);
367 __bcache_io_error(c, "btree IO error %i", ret);
372 /* If some a bucket wasn't written, we can't erasure code it: */
373 for_each_set_bit(dev, op->failed.d, BCH_SB_MEMBERS_MAX)
374 bch2_open_bucket_write_error(c, &op->open_buckets, dev);
376 bch2_open_buckets_put(c, &op->open_buckets);
379 keys->top = keys->keys;
384 static void bch2_write_index(struct closure *cl)
386 struct bch_write_op *op = container_of(cl, struct bch_write_op, cl);
387 struct bch_fs *c = op->c;
389 __bch2_write_index(op);
391 if (!op->error && (op->flags & BCH_WRITE_FLUSH)) {
392 bch2_journal_flush_seq_async(&c->journal,
395 continue_at(cl, bch2_write_done, index_update_wq(op));
397 continue_at_nobarrier(cl, bch2_write_done, NULL);
401 static void bch2_write_endio(struct bio *bio)
403 struct closure *cl = bio->bi_private;
404 struct bch_write_op *op = container_of(cl, struct bch_write_op, cl);
405 struct bch_write_bio *wbio = to_wbio(bio);
406 struct bch_write_bio *parent = wbio->split ? wbio->parent : NULL;
407 struct bch_fs *c = wbio->c;
408 struct bch_dev *ca = bch_dev_bkey_exists(c, wbio->dev);
410 if (bch2_dev_io_err_on(bio->bi_status, ca, "data write"))
411 set_bit(wbio->dev, op->failed.d);
413 if (wbio->have_ioref) {
414 bch2_latency_acct(ca, wbio->submit_time, WRITE);
415 percpu_ref_put(&ca->io_ref);
419 bch2_bio_free_pages_pool(c, bio);
425 bio_endio(&parent->bio);
430 static void init_append_extent(struct bch_write_op *op,
431 struct write_point *wp,
432 struct bversion version,
433 struct bch_extent_crc_unpacked crc)
435 struct bkey_i_extent *e = bkey_extent_init(op->insert_keys.top);
436 struct bch_extent_ptr *ptr;
438 op->pos.offset += crc.uncompressed_size;
440 e->k.size = crc.uncompressed_size;
441 e->k.version = version;
443 bch2_extent_crc_append(e, crc);
444 bch2_alloc_sectors_append_ptrs(op->c, wp, &e->k_i,
445 crc.compressed_size);
447 if (op->flags & BCH_WRITE_CACHED)
448 extent_for_each_ptr(extent_i_to_s(e), ptr)
451 bch2_keylist_push(&op->insert_keys);
454 static struct bio *bch2_write_bio_alloc(struct bch_fs *c,
455 struct write_point *wp,
457 bool *page_alloc_failed,
460 struct bch_write_bio *wbio;
462 unsigned output_available =
463 min(wp->sectors_free << 9, src->bi_iter.bi_size);
464 unsigned pages = DIV_ROUND_UP(output_available, PAGE_SIZE);
466 bio = bio_alloc_bioset(GFP_NOIO, pages, &c->bio_write);
467 wbio = wbio_init(bio);
468 wbio->put_bio = true;
469 /* copy WRITE_SYNC flag */
470 wbio->bio.bi_opf = src->bi_opf;
473 bio->bi_iter.bi_size = output_available;
474 bch2_bio_map(bio, buf);
481 * We can't use mempool for more than c->sb.encoded_extent_max
482 * worth of pages, but we'd like to allocate more if we can:
484 while (bio->bi_iter.bi_size < output_available) {
485 unsigned len = min_t(unsigned, PAGE_SIZE,
486 output_available - bio->bi_iter.bi_size);
489 p = alloc_page(GFP_NOIO);
492 min_t(unsigned, output_available,
493 c->sb.encoded_extent_max << 9);
495 if (bio_sectors(bio) < pool_max)
496 bch2_bio_alloc_pages_pool(c, bio, pool_max);
500 bio->bi_io_vec[bio->bi_vcnt++] = (struct bio_vec) {
505 bio->bi_iter.bi_size += len;
508 *page_alloc_failed = bio->bi_vcnt < pages;
512 static int bch2_write_rechecksum(struct bch_fs *c,
513 struct bch_write_op *op,
514 unsigned new_csum_type)
516 struct bio *bio = &op->wbio.bio;
517 struct bch_extent_crc_unpacked new_crc;
520 /* bch2_rechecksum_bio() can't encrypt or decrypt data: */
522 if (bch2_csum_type_is_encryption(op->crc.csum_type) !=
523 bch2_csum_type_is_encryption(new_csum_type))
524 new_csum_type = op->crc.csum_type;
526 ret = bch2_rechecksum_bio(c, bio, op->version, op->crc,
528 op->crc.offset, op->crc.live_size,
533 bio_advance(bio, op->crc.offset << 9);
534 bio->bi_iter.bi_size = op->crc.live_size << 9;
539 static int bch2_write_decrypt(struct bch_write_op *op)
541 struct bch_fs *c = op->c;
542 struct nonce nonce = extent_nonce(op->version, op->crc);
543 struct bch_csum csum;
545 if (!bch2_csum_type_is_encryption(op->crc.csum_type))
549 * If we need to decrypt data in the write path, we'll no longer be able
550 * to verify the existing checksum (poly1305 mac, in this case) after
551 * it's decrypted - this is the last point we'll be able to reverify the
554 csum = bch2_checksum_bio(c, op->crc.csum_type, nonce, &op->wbio.bio);
555 if (bch2_crc_cmp(op->crc.csum, csum))
558 bch2_encrypt_bio(c, op->crc.csum_type, nonce, &op->wbio.bio);
559 op->crc.csum_type = 0;
560 op->crc.csum = (struct bch_csum) { 0, 0 };
564 static enum prep_encoded_ret {
567 PREP_ENCODED_CHECKSUM_ERR,
568 PREP_ENCODED_DO_WRITE,
569 } bch2_write_prep_encoded_data(struct bch_write_op *op, struct write_point *wp)
571 struct bch_fs *c = op->c;
572 struct bio *bio = &op->wbio.bio;
574 if (!(op->flags & BCH_WRITE_DATA_ENCODED))
575 return PREP_ENCODED_OK;
577 BUG_ON(bio_sectors(bio) != op->crc.compressed_size);
579 /* Can we just write the entire extent as is? */
580 if (op->crc.uncompressed_size == op->crc.live_size &&
581 op->crc.compressed_size <= wp->sectors_free &&
582 op->crc.compression_type == op->compression_type) {
583 if (!op->crc.compression_type &&
584 op->csum_type != op->crc.csum_type &&
585 bch2_write_rechecksum(c, op, op->csum_type))
586 return PREP_ENCODED_CHECKSUM_ERR;
588 return PREP_ENCODED_DO_WRITE;
592 * If the data is compressed and we couldn't write the entire extent as
593 * is, we have to decompress it:
595 if (op->crc.compression_type) {
596 struct bch_csum csum;
598 if (bch2_write_decrypt(op))
599 return PREP_ENCODED_CHECKSUM_ERR;
601 /* Last point we can still verify checksum: */
602 csum = bch2_checksum_bio(c, op->crc.csum_type,
603 extent_nonce(op->version, op->crc),
605 if (bch2_crc_cmp(op->crc.csum, csum))
606 return PREP_ENCODED_CHECKSUM_ERR;
608 if (bch2_bio_uncompress_inplace(c, bio, &op->crc))
609 return PREP_ENCODED_ERR;
613 * No longer have compressed data after this point - data might be
618 * If the data is checksummed and we're only writing a subset,
619 * rechecksum and adjust bio to point to currently live data:
621 if ((op->crc.live_size != op->crc.uncompressed_size ||
622 op->crc.csum_type != op->csum_type) &&
623 bch2_write_rechecksum(c, op, op->csum_type))
624 return PREP_ENCODED_CHECKSUM_ERR;
627 * If we want to compress the data, it has to be decrypted:
629 if ((op->compression_type ||
630 bch2_csum_type_is_encryption(op->crc.csum_type) !=
631 bch2_csum_type_is_encryption(op->csum_type)) &&
632 bch2_write_decrypt(op))
633 return PREP_ENCODED_CHECKSUM_ERR;
635 return PREP_ENCODED_OK;
638 static int bch2_write_extent(struct bch_write_op *op, struct write_point *wp)
640 struct bch_fs *c = op->c;
641 struct bio *src = &op->wbio.bio, *dst = src;
642 struct bvec_iter saved_iter;
643 struct bkey_i *key_to_write;
645 unsigned key_to_write_offset = op->insert_keys.top_p -
646 op->insert_keys.keys_p;
647 unsigned total_output = 0, total_input = 0;
649 bool page_alloc_failed = false;
652 BUG_ON(!bio_sectors(src));
654 ec_buf = bch2_writepoint_ec_buf(c, wp);
656 switch (bch2_write_prep_encoded_data(op, wp)) {
657 case PREP_ENCODED_OK:
659 case PREP_ENCODED_ERR:
662 case PREP_ENCODED_CHECKSUM_ERR:
664 case PREP_ENCODED_DO_WRITE:
666 dst = bch2_write_bio_alloc(c, wp, src,
669 bio_copy_data(dst, src);
672 init_append_extent(op, wp, op->version, op->crc);
677 op->compression_type ||
679 !(op->flags & BCH_WRITE_PAGES_STABLE)) ||
680 (bch2_csum_type_is_encryption(op->csum_type) &&
681 !(op->flags & BCH_WRITE_PAGES_OWNED))) {
682 dst = bch2_write_bio_alloc(c, wp, src,
688 saved_iter = dst->bi_iter;
691 struct bch_extent_crc_unpacked crc =
692 (struct bch_extent_crc_unpacked) { 0 };
693 struct bversion version = op->version;
694 size_t dst_len, src_len;
696 if (page_alloc_failed &&
697 bio_sectors(dst) < wp->sectors_free &&
698 bio_sectors(dst) < c->sb.encoded_extent_max)
701 BUG_ON(op->compression_type &&
702 (op->flags & BCH_WRITE_DATA_ENCODED) &&
703 bch2_csum_type_is_encryption(op->crc.csum_type));
704 BUG_ON(op->compression_type && !bounce);
706 crc.compression_type = op->compression_type
707 ? bch2_bio_compress(c, dst, &dst_len, src, &src_len,
708 op->compression_type)
710 if (!crc.compression_type) {
711 dst_len = min(dst->bi_iter.bi_size, src->bi_iter.bi_size);
712 dst_len = min_t(unsigned, dst_len, wp->sectors_free << 9);
715 dst_len = min_t(unsigned, dst_len,
716 c->sb.encoded_extent_max << 9);
719 swap(dst->bi_iter.bi_size, dst_len);
720 bio_copy_data(dst, src);
721 swap(dst->bi_iter.bi_size, dst_len);
727 BUG_ON(!src_len || !dst_len);
729 if (bch2_csum_type_is_encryption(op->csum_type)) {
730 if (bversion_zero(version)) {
731 version.lo = atomic64_inc_return(&c->key_version) + 1;
733 crc.nonce = op->nonce;
734 op->nonce += src_len >> 9;
738 if ((op->flags & BCH_WRITE_DATA_ENCODED) &&
739 !crc.compression_type &&
740 bch2_csum_type_is_encryption(op->crc.csum_type) ==
741 bch2_csum_type_is_encryption(op->csum_type)) {
743 * Note: when we're using rechecksum(), we need to be
744 * checksumming @src because it has all the data our
745 * existing checksum covers - if we bounced (because we
746 * were trying to compress), @dst will only have the
747 * part of the data the new checksum will cover.
749 * But normally we want to be checksumming post bounce,
750 * because part of the reason for bouncing is so the
751 * data can't be modified (by userspace) while it's in
754 if (bch2_rechecksum_bio(c, src, version, op->crc,
757 bio_sectors(src) - (src_len >> 9),
761 if ((op->flags & BCH_WRITE_DATA_ENCODED) &&
762 bch2_rechecksum_bio(c, src, version, op->crc,
765 bio_sectors(src) - (src_len >> 9),
769 crc.compressed_size = dst_len >> 9;
770 crc.uncompressed_size = src_len >> 9;
771 crc.live_size = src_len >> 9;
773 swap(dst->bi_iter.bi_size, dst_len);
774 bch2_encrypt_bio(c, op->csum_type,
775 extent_nonce(version, crc), dst);
776 crc.csum = bch2_checksum_bio(c, op->csum_type,
777 extent_nonce(version, crc), dst);
778 crc.csum_type = op->csum_type;
779 swap(dst->bi_iter.bi_size, dst_len);
782 init_append_extent(op, wp, version, crc);
785 bio_advance(dst, dst_len);
786 bio_advance(src, src_len);
787 total_output += dst_len;
788 total_input += src_len;
789 } while (dst->bi_iter.bi_size &&
790 src->bi_iter.bi_size &&
792 !bch2_keylist_realloc(&op->insert_keys,
794 ARRAY_SIZE(op->inline_keys),
795 BKEY_EXTENT_U64s_MAX));
797 more = src->bi_iter.bi_size != 0;
799 dst->bi_iter = saved_iter;
801 if (dst == src && more) {
802 BUG_ON(total_output != total_input);
804 dst = bio_split(src, total_input >> 9,
805 GFP_NOIO, &c->bio_write);
806 wbio_init(dst)->put_bio = true;
807 /* copy WRITE_SYNC flag */
808 dst->bi_opf = src->bi_opf;
811 dst->bi_iter.bi_size = total_output;
813 /* Free unneeded pages after compressing: */
814 if (to_wbio(dst)->bounce)
815 while (dst->bi_vcnt > DIV_ROUND_UP(dst->bi_iter.bi_size, PAGE_SIZE))
816 mempool_free(dst->bi_io_vec[--dst->bi_vcnt].bv_page,
817 &c->bio_bounce_pages);
819 /* might have done a realloc... */
821 key_to_write = (void *) (op->insert_keys.keys_p + key_to_write_offset);
823 bch2_ec_add_backpointer(c, wp,
824 bkey_start_pos(&key_to_write->k),
827 dst->bi_end_io = bch2_write_endio;
828 dst->bi_private = &op->cl;
829 bio_set_op_attrs(dst, REQ_OP_WRITE, 0);
831 closure_get(dst->bi_private);
833 bch2_submit_wbio_replicas(to_wbio(dst), c, BCH_DATA_USER,
837 bch_err(c, "error verifying existing checksum while "
838 "rewriting existing data (memory corruption?)");
841 if (to_wbio(dst)->bounce)
842 bch2_bio_free_pages_pool(c, dst);
843 if (to_wbio(dst)->put_bio)
849 static void __bch2_write(struct closure *cl)
851 struct bch_write_op *op = container_of(cl, struct bch_write_op, cl);
852 struct bch_fs *c = op->c;
853 struct write_point *wp;
856 memset(&op->failed, 0, sizeof(op->failed));
859 /* +1 for possible cache device: */
860 if (op->open_buckets.nr + op->nr_replicas + 1 >
861 ARRAY_SIZE(op->open_buckets.v))
864 if (bch2_keylist_realloc(&op->insert_keys,
866 ARRAY_SIZE(op->inline_keys),
867 BKEY_EXTENT_U64s_MAX))
870 wp = bch2_alloc_sectors_start(c,
872 op->opts.erasure_code,
876 op->nr_replicas_required,
879 (op->flags & BCH_WRITE_ALLOC_NOWAIT) ? NULL : cl);
882 if (unlikely(IS_ERR(wp))) {
883 if (unlikely(PTR_ERR(wp) != -EAGAIN)) {
891 ret = bch2_write_extent(op, wp);
893 bch2_open_bucket_get(c, wp, &op->open_buckets);
894 bch2_alloc_sectors_done(c, wp);
900 continue_at(cl, bch2_write_index, index_update_wq(op));
905 continue_at(cl, !bch2_keylist_empty(&op->insert_keys)
907 : bch2_write_done, index_update_wq(op));
912 if (!bch2_keylist_empty(&op->insert_keys)) {
913 __bch2_write_index(op);
916 continue_at_nobarrier(cl, bch2_write_done, NULL);
925 * bch_write - handle a write to a cache device or flash only volume
927 * This is the starting point for any data to end up in a cache device; it could
928 * be from a normal write, or a writeback write, or a write to a flash only
929 * volume - it's also used by the moving garbage collector to compact data in
930 * mostly empty buckets.
932 * It first writes the data to the cache, creating a list of keys to be inserted
933 * (if the data won't fit in a single open bucket, there will be multiple keys);
934 * after the data is written it calls bch_journal, and after the keys have been
935 * added to the next journal write they're inserted into the btree.
937 * If op->discard is true, instead of inserting the data it invalidates the
938 * region of the cache represented by op->bio and op->inode.
940 void bch2_write(struct closure *cl)
942 struct bch_write_op *op = container_of(cl, struct bch_write_op, cl);
943 struct bch_fs *c = op->c;
945 BUG_ON(!op->nr_replicas);
946 BUG_ON(!op->write_point.v);
947 BUG_ON(!bkey_cmp(op->pos, POS_MAX));
948 BUG_ON(bio_sectors(&op->wbio.bio) > U16_MAX);
950 op->start_time = local_clock();
952 bch2_keylist_init(&op->insert_keys, op->inline_keys);
953 wbio_init(&op->wbio.bio)->put_bio = false;
955 if (c->opts.nochanges ||
956 !percpu_ref_tryget(&c->writes)) {
957 __bcache_io_error(c, "read only");
959 if (!(op->flags & BCH_WRITE_NOPUT_RESERVATION))
960 bch2_disk_reservation_put(c, &op->res);
965 bch2_increment_clock(c, bio_sectors(&op->wbio.bio), WRITE);
967 continue_at_nobarrier(cl, __bch2_write, NULL);
970 /* Cache promotion on read */
977 struct rhash_head hash;
980 struct migrate_write write;
981 struct bio_vec bi_inline_vecs[0]; /* must be last */
984 static const struct rhashtable_params bch_promote_params = {
985 .head_offset = offsetof(struct promote_op, hash),
986 .key_offset = offsetof(struct promote_op, pos),
987 .key_len = sizeof(struct bpos),
990 static inline bool should_promote(struct bch_fs *c, struct bkey_s_c k,
992 struct bch_io_opts opts,
995 if (!opts.promote_target)
998 if (!(flags & BCH_READ_MAY_PROMOTE))
1001 if (percpu_ref_is_dying(&c->writes))
1004 if (!bkey_extent_is_data(k.k))
1007 if (bch2_extent_has_target(c, bkey_s_c_to_extent(k), opts.promote_target))
1010 if (bch2_target_congested(c, opts.promote_target))
1013 if (rhashtable_lookup_fast(&c->promote_table, &pos,
1014 bch_promote_params))
1020 static void promote_free(struct bch_fs *c, struct promote_op *op)
1024 ret = rhashtable_remove_fast(&c->promote_table, &op->hash,
1025 bch_promote_params);
1027 percpu_ref_put(&c->writes);
1031 static void promote_done(struct closure *cl)
1033 struct promote_op *op =
1034 container_of(cl, struct promote_op, cl);
1035 struct bch_fs *c = op->write.op.c;
1037 bch2_time_stats_update(&c->times[BCH_TIME_data_promote],
1040 bch2_bio_free_pages_pool(c, &op->write.op.wbio.bio);
1041 promote_free(c, op);
1044 static void promote_start(struct promote_op *op, struct bch_read_bio *rbio)
1046 struct bch_fs *c = rbio->c;
1047 struct closure *cl = &op->cl;
1048 struct bio *bio = &op->write.op.wbio.bio;
1050 trace_promote(&rbio->bio);
1052 /* we now own pages: */
1053 BUG_ON(!rbio->bounce);
1054 BUG_ON(rbio->bio.bi_vcnt > bio->bi_max_vecs);
1056 memcpy(bio->bi_io_vec, rbio->bio.bi_io_vec,
1057 sizeof(struct bio_vec) * rbio->bio.bi_vcnt);
1058 swap(bio->bi_vcnt, rbio->bio.bi_vcnt);
1060 bch2_migrate_read_done(&op->write, rbio);
1062 closure_init(cl, NULL);
1063 closure_call(&op->write.op.cl, bch2_write, c->wq, cl);
1064 closure_return_with_destructor(cl, promote_done);
1068 static struct promote_op *__promote_alloc(struct bch_fs *c,
1070 struct extent_ptr_decoded *pick,
1071 struct bch_io_opts opts,
1072 unsigned rbio_sectors,
1073 struct bch_read_bio **rbio)
1075 struct promote_op *op = NULL;
1077 unsigned rbio_pages = DIV_ROUND_UP(rbio_sectors, PAGE_SECTORS);
1078 /* data might have to be decompressed in the write path: */
1079 unsigned wbio_pages = DIV_ROUND_UP(pick->crc.uncompressed_size,
1083 if (!percpu_ref_tryget(&c->writes))
1086 op = kzalloc(sizeof(*op) + sizeof(struct bio_vec) * wbio_pages,
1091 op->start_time = local_clock();
1095 * promotes require bouncing, but if the extent isn't
1096 * checksummed/compressed it might be too big for the mempool:
1098 if (rbio_sectors > c->sb.encoded_extent_max) {
1099 *rbio = kzalloc(sizeof(struct bch_read_bio) +
1100 sizeof(struct bio_vec) * rbio_pages,
1105 rbio_init(&(*rbio)->bio, opts);
1106 bio_init(&(*rbio)->bio, (*rbio)->bio.bi_inline_vecs,
1109 (*rbio)->bio.bi_iter.bi_size = rbio_sectors << 9;
1110 bch2_bio_map(&(*rbio)->bio, NULL);
1112 if (bch2_bio_alloc_pages(&(*rbio)->bio, GFP_NOIO))
1115 (*rbio)->bounce = true;
1116 (*rbio)->split = true;
1117 (*rbio)->kmalloc = true;
1120 if (rhashtable_lookup_insert_fast(&c->promote_table, &op->hash,
1121 bch_promote_params))
1124 bio = &op->write.op.wbio.bio;
1125 bio_init(bio, bio->bi_inline_vecs, wbio_pages);
1127 ret = bch2_migrate_write_init(c, &op->write,
1128 writepoint_hashed((unsigned long) current),
1131 (struct data_opts) {
1132 .target = opts.promote_target
1140 bio_free_pages(&(*rbio)->bio);
1144 percpu_ref_put(&c->writes);
1148 static inline struct promote_op *promote_alloc(struct bch_fs *c,
1149 struct bvec_iter iter,
1151 struct extent_ptr_decoded *pick,
1152 struct bch_io_opts opts,
1154 struct bch_read_bio **rbio,
1158 bool promote_full = *read_full || READ_ONCE(c->promote_whole_extents);
1159 unsigned sectors = promote_full
1160 ? pick->crc.compressed_size
1161 : bvec_iter_sectors(iter);
1162 struct bpos pos = promote_full
1163 ? bkey_start_pos(k.k)
1164 : POS(k.k->p.inode, iter.bi_sector);
1165 struct promote_op *promote;
1167 if (!should_promote(c, k, pos, opts, flags))
1170 promote = __promote_alloc(c, pos, pick, opts, sectors, rbio);
1175 *read_full = promote_full;
1181 #define READ_RETRY_AVOID 1
1182 #define READ_RETRY 2
1187 RBIO_CONTEXT_HIGHPRI,
1188 RBIO_CONTEXT_UNBOUND,
1191 static inline struct bch_read_bio *
1192 bch2_rbio_parent(struct bch_read_bio *rbio)
1194 return rbio->split ? rbio->parent : rbio;
1198 static void bch2_rbio_punt(struct bch_read_bio *rbio, work_func_t fn,
1199 enum rbio_context context,
1200 struct workqueue_struct *wq)
1202 if (context <= rbio->context) {
1205 rbio->work.func = fn;
1206 rbio->context = context;
1207 queue_work(wq, &rbio->work);
1211 static inline struct bch_read_bio *bch2_rbio_free(struct bch_read_bio *rbio)
1213 BUG_ON(rbio->bounce && !rbio->split);
1216 promote_free(rbio->c, rbio->promote);
1217 rbio->promote = NULL;
1220 bch2_bio_free_pages_pool(rbio->c, &rbio->bio);
1223 struct bch_read_bio *parent = rbio->parent;
1228 bio_put(&rbio->bio);
1236 static void bch2_rbio_done(struct bch_read_bio *rbio)
1238 bch2_time_stats_update(&rbio->c->times[BCH_TIME_data_read],
1240 bio_endio(&rbio->bio);
1243 static void bch2_read_retry_nodecode(struct bch_fs *c, struct bch_read_bio *rbio,
1244 struct bvec_iter bvec_iter, u64 inode,
1245 struct bch_io_failures *failed,
1248 struct btree_trans trans;
1249 struct btree_iter *iter;
1254 flags &= ~BCH_READ_LAST_FRAGMENT;
1256 bch2_trans_init(&trans, c);
1258 iter = bch2_trans_get_iter(&trans, BTREE_ID_EXTENTS,
1259 rbio->pos, BTREE_ITER_SLOTS);
1261 rbio->bio.bi_status = 0;
1263 k = bch2_btree_iter_peek_slot(iter);
1267 bkey_reassemble(&tmp.k, k);
1268 k = bkey_i_to_s_c(&tmp.k);
1269 bch2_trans_unlock(&trans);
1271 if (!bkey_extent_is_data(k.k) ||
1272 !bch2_extent_matches_ptr(c, bkey_i_to_s_c_extent(&tmp.k),
1275 rbio->pick.crc.offset)) {
1276 /* extent we wanted to read no longer exists: */
1281 ret = __bch2_read_extent(c, rbio, bvec_iter, k, failed, flags);
1282 if (ret == READ_RETRY)
1287 bch2_rbio_done(rbio);
1288 bch2_trans_exit(&trans);
1291 rbio->bio.bi_status = BLK_STS_IOERR;
1295 static void bch2_read_retry(struct bch_fs *c, struct bch_read_bio *rbio,
1296 struct bvec_iter bvec_iter, u64 inode,
1297 struct bch_io_failures *failed, unsigned flags)
1299 struct btree_trans trans;
1300 struct btree_iter *iter;
1304 bch2_trans_init(&trans, c);
1306 flags &= ~BCH_READ_LAST_FRAGMENT;
1307 flags |= BCH_READ_MUST_CLONE;
1309 for_each_btree_key(&trans, iter, BTREE_ID_EXTENTS,
1310 POS(inode, bvec_iter.bi_sector),
1311 BTREE_ITER_SLOTS, k, ret) {
1315 bkey_reassemble(&tmp.k, k);
1316 k = bkey_i_to_s_c(&tmp.k);
1317 bch2_btree_trans_unlock(&trans);
1319 bytes = min_t(unsigned, bvec_iter.bi_size,
1320 (k.k->p.offset - bvec_iter.bi_sector) << 9);
1321 swap(bvec_iter.bi_size, bytes);
1323 ret = __bch2_read_extent(c, rbio, bvec_iter, k, failed, flags);
1331 if (bytes == bvec_iter.bi_size)
1334 swap(bvec_iter.bi_size, bytes);
1335 bio_advance_iter(&rbio->bio, &bvec_iter, bytes);
1339 * If we get here, it better have been because there was an error
1340 * reading a btree node
1343 __bcache_io_error(c, "btree IO error: %i", ret);
1345 rbio->bio.bi_status = BLK_STS_IOERR;
1347 bch2_trans_exit(&trans);
1348 bch2_rbio_done(rbio);
1351 static void bch2_rbio_retry(struct work_struct *work)
1353 struct bch_read_bio *rbio =
1354 container_of(work, struct bch_read_bio, work);
1355 struct bch_fs *c = rbio->c;
1356 struct bvec_iter iter = rbio->bvec_iter;
1357 unsigned flags = rbio->flags;
1358 u64 inode = rbio->pos.inode;
1359 struct bch_io_failures failed = { .nr = 0 };
1361 trace_read_retry(&rbio->bio);
1363 if (rbio->retry == READ_RETRY_AVOID)
1364 bch2_mark_io_failure(&failed, &rbio->pick);
1366 rbio->bio.bi_status = 0;
1368 rbio = bch2_rbio_free(rbio);
1370 flags |= BCH_READ_IN_RETRY;
1371 flags &= ~BCH_READ_MAY_PROMOTE;
1373 if (flags & BCH_READ_NODECODE)
1374 bch2_read_retry_nodecode(c, rbio, iter, inode, &failed, flags);
1376 bch2_read_retry(c, rbio, iter, inode, &failed, flags);
1379 static void bch2_rbio_error(struct bch_read_bio *rbio, int retry,
1382 rbio->retry = retry;
1384 if (rbio->flags & BCH_READ_IN_RETRY)
1387 if (retry == READ_ERR) {
1388 rbio = bch2_rbio_free(rbio);
1390 rbio->bio.bi_status = error;
1391 bch2_rbio_done(rbio);
1393 bch2_rbio_punt(rbio, bch2_rbio_retry,
1394 RBIO_CONTEXT_UNBOUND, system_unbound_wq);
1398 static void bch2_rbio_narrow_crcs(struct bch_read_bio *rbio)
1400 struct bch_fs *c = rbio->c;
1401 struct btree_trans trans;
1402 struct btree_iter *iter;
1404 struct bkey_i_extent *e;
1406 struct bch_extent_crc_unpacked new_crc;
1410 if (rbio->pick.crc.compression_type)
1413 bch2_trans_init(&trans, c);
1415 bch2_trans_begin(&trans);
1417 iter = bch2_trans_get_iter(&trans, BTREE_ID_EXTENTS, rbio->pos,
1419 k = bch2_btree_iter_peek(iter);
1420 if (IS_ERR_OR_NULL(k.k))
1423 if (!bkey_extent_is_data(k.k))
1426 bkey_reassemble(&new.k, k);
1427 e = bkey_i_to_extent(&new.k);
1429 if (!bch2_extent_matches_ptr(c, extent_i_to_s_c(e),
1432 rbio->pick.crc.offset) ||
1433 bversion_cmp(e->k.version, rbio->version))
1436 /* Extent was merged? */
1437 if (bkey_start_offset(&e->k) < rbio->pos.offset ||
1438 e->k.p.offset > rbio->pos.offset + rbio->pick.crc.uncompressed_size)
1441 /* The extent might have been partially overwritten since we read it: */
1442 offset = rbio->pick.crc.offset + (bkey_start_offset(&e->k) - rbio->pos.offset);
1444 if (bch2_rechecksum_bio(c, &rbio->bio, rbio->version,
1445 rbio->pick.crc, NULL, &new_crc,
1447 rbio->pick.crc.csum_type)) {
1448 bch_err(c, "error verifying existing checksum while narrowing checksum (memory corruption?)");
1452 if (!bch2_extent_narrow_crcs(e, new_crc))
1455 bch2_trans_update(&trans, BTREE_INSERT_ENTRY(iter, &e->k_i));
1456 ret = bch2_trans_commit(&trans, NULL, NULL,
1457 BTREE_INSERT_ATOMIC|
1458 BTREE_INSERT_NOFAIL|
1459 BTREE_INSERT_NOWAIT);
1463 bch2_trans_exit(&trans);
1466 static bool should_narrow_crcs(struct bkey_s_c k,
1467 struct extent_ptr_decoded *pick,
1470 return !(flags & BCH_READ_IN_RETRY) &&
1471 bkey_extent_is_data(k.k) &&
1472 bch2_can_narrow_extent_crcs(bkey_s_c_to_extent(k), pick->crc);
1475 /* Inner part that may run in process context */
1476 static void __bch2_read_endio(struct work_struct *work)
1478 struct bch_read_bio *rbio =
1479 container_of(work, struct bch_read_bio, work);
1480 struct bch_fs *c = rbio->c;
1481 struct bch_dev *ca = bch_dev_bkey_exists(c, rbio->pick.ptr.dev);
1482 struct bio *src = &rbio->bio;
1483 struct bio *dst = &bch2_rbio_parent(rbio)->bio;
1484 struct bvec_iter dst_iter = rbio->bvec_iter;
1485 struct bch_extent_crc_unpacked crc = rbio->pick.crc;
1486 struct nonce nonce = extent_nonce(rbio->version, crc);
1487 struct bch_csum csum;
1489 /* Reset iterator for checksumming and copying bounced data: */
1491 src->bi_iter.bi_size = crc.compressed_size << 9;
1492 src->bi_iter.bi_idx = 0;
1493 src->bi_iter.bi_bvec_done = 0;
1495 src->bi_iter = rbio->bvec_iter;
1498 csum = bch2_checksum_bio(c, crc.csum_type, nonce, src);
1499 if (bch2_crc_cmp(csum, rbio->pick.crc.csum))
1502 if (unlikely(rbio->narrow_crcs))
1503 bch2_rbio_narrow_crcs(rbio);
1505 if (rbio->flags & BCH_READ_NODECODE)
1508 /* Adjust crc to point to subset of data we want: */
1509 crc.offset += rbio->bvec_iter.bi_sector - rbio->pos.offset;
1510 crc.live_size = bvec_iter_sectors(rbio->bvec_iter);
1512 if (crc.compression_type != BCH_COMPRESSION_NONE) {
1513 bch2_encrypt_bio(c, crc.csum_type, nonce, src);
1514 if (bch2_bio_uncompress(c, src, dst, dst_iter, crc))
1515 goto decompression_err;
1517 /* don't need to decrypt the entire bio: */
1518 nonce = nonce_add(nonce, crc.offset << 9);
1519 bio_advance(src, crc.offset << 9);
1521 BUG_ON(src->bi_iter.bi_size < dst_iter.bi_size);
1522 src->bi_iter.bi_size = dst_iter.bi_size;
1524 bch2_encrypt_bio(c, crc.csum_type, nonce, src);
1527 struct bvec_iter src_iter = src->bi_iter;
1528 bio_copy_data_iter(dst, &dst_iter, src, &src_iter);
1532 if (rbio->promote) {
1534 * Re encrypt data we decrypted, so it's consistent with
1537 bch2_encrypt_bio(c, crc.csum_type, nonce, src);
1538 promote_start(rbio->promote, rbio);
1539 rbio->promote = NULL;
1542 if (likely(!(rbio->flags & BCH_READ_IN_RETRY))) {
1543 rbio = bch2_rbio_free(rbio);
1544 bch2_rbio_done(rbio);
1549 * Checksum error: if the bio wasn't bounced, we may have been
1550 * reading into buffers owned by userspace (that userspace can
1551 * scribble over) - retry the read, bouncing it this time:
1553 if (!rbio->bounce && (rbio->flags & BCH_READ_USER_MAPPED)) {
1554 rbio->flags |= BCH_READ_MUST_BOUNCE;
1555 bch2_rbio_error(rbio, READ_RETRY, BLK_STS_IOERR);
1559 bch2_dev_io_error(ca,
1560 "data checksum error, inode %llu offset %llu: expected %0llx:%0llx got %0llx:%0llx (type %u)",
1561 rbio->pos.inode, (u64) rbio->bvec_iter.bi_sector,
1562 rbio->pick.crc.csum.hi, rbio->pick.crc.csum.lo,
1563 csum.hi, csum.lo, crc.csum_type);
1564 bch2_rbio_error(rbio, READ_RETRY_AVOID, BLK_STS_IOERR);
1567 __bcache_io_error(c, "decompression error, inode %llu offset %llu",
1569 (u64) rbio->bvec_iter.bi_sector);
1570 bch2_rbio_error(rbio, READ_ERR, BLK_STS_IOERR);
1574 static void bch2_read_endio(struct bio *bio)
1576 struct bch_read_bio *rbio =
1577 container_of(bio, struct bch_read_bio, bio);
1578 struct bch_fs *c = rbio->c;
1579 struct bch_dev *ca = bch_dev_bkey_exists(c, rbio->pick.ptr.dev);
1580 struct workqueue_struct *wq = NULL;
1581 enum rbio_context context = RBIO_CONTEXT_NULL;
1583 if (rbio->have_ioref) {
1584 bch2_latency_acct(ca, rbio->submit_time, READ);
1585 percpu_ref_put(&ca->io_ref);
1589 rbio->bio.bi_end_io = rbio->end_io;
1591 if (bch2_dev_io_err_on(bio->bi_status, ca, "data read")) {
1592 bch2_rbio_error(rbio, READ_RETRY_AVOID, bio->bi_status);
1596 if (rbio->pick.ptr.cached &&
1597 (((rbio->flags & BCH_READ_RETRY_IF_STALE) && race_fault()) ||
1598 ptr_stale(ca, &rbio->pick.ptr))) {
1599 atomic_long_inc(&c->read_realloc_races);
1601 if (rbio->flags & BCH_READ_RETRY_IF_STALE)
1602 bch2_rbio_error(rbio, READ_RETRY, BLK_STS_AGAIN);
1604 bch2_rbio_error(rbio, READ_ERR, BLK_STS_AGAIN);
1608 if (rbio->narrow_crcs ||
1609 rbio->pick.crc.compression_type ||
1610 bch2_csum_type_is_encryption(rbio->pick.crc.csum_type))
1611 context = RBIO_CONTEXT_UNBOUND, wq = system_unbound_wq;
1612 else if (rbio->pick.crc.csum_type)
1613 context = RBIO_CONTEXT_HIGHPRI, wq = system_highpri_wq;
1615 bch2_rbio_punt(rbio, __bch2_read_endio, context, wq);
1618 int __bch2_read_extent(struct bch_fs *c, struct bch_read_bio *orig,
1619 struct bvec_iter iter, struct bkey_s_c k,
1620 struct bch_io_failures *failed, unsigned flags)
1622 struct extent_ptr_decoded pick;
1623 struct bch_read_bio *rbio = NULL;
1625 struct promote_op *promote = NULL;
1626 bool bounce = false, read_full = false, narrow_crcs = false;
1627 struct bpos pos = bkey_start_pos(k.k);
1630 pick_ret = bch2_bkey_pick_read_device(c, k, failed, &pick);
1632 /* hole or reservation - just zero fill: */
1637 __bcache_io_error(c, "no device to read from");
1642 ca = bch_dev_bkey_exists(c, pick.ptr.dev);
1644 if (flags & BCH_READ_NODECODE) {
1646 * can happen if we retry, and the extent we were going to read
1647 * has been merged in the meantime:
1649 if (pick.crc.compressed_size > orig->bio.bi_vcnt * PAGE_SECTORS)
1652 iter.bi_sector = pos.offset;
1653 iter.bi_size = pick.crc.compressed_size << 9;
1657 if (!(flags & BCH_READ_LAST_FRAGMENT) ||
1658 bio_flagged(&orig->bio, BIO_CHAIN))
1659 flags |= BCH_READ_MUST_CLONE;
1661 narrow_crcs = should_narrow_crcs(k, &pick, flags);
1663 if (narrow_crcs && (flags & BCH_READ_USER_MAPPED))
1664 flags |= BCH_READ_MUST_BOUNCE;
1666 EBUG_ON(bkey_start_offset(k.k) > iter.bi_sector ||
1667 k.k->p.offset < bvec_iter_end_sector(iter));
1669 if (pick.crc.compression_type != BCH_COMPRESSION_NONE ||
1670 (pick.crc.csum_type != BCH_CSUM_NONE &&
1671 (bvec_iter_sectors(iter) != pick.crc.uncompressed_size ||
1672 (bch2_csum_type_is_encryption(pick.crc.csum_type) &&
1673 (flags & BCH_READ_USER_MAPPED)) ||
1674 (flags & BCH_READ_MUST_BOUNCE)))) {
1679 promote = promote_alloc(c, iter, k, &pick, orig->opts, flags,
1680 &rbio, &bounce, &read_full);
1683 EBUG_ON(pick.crc.compression_type);
1684 EBUG_ON(pick.crc.csum_type &&
1685 (bvec_iter_sectors(iter) != pick.crc.uncompressed_size ||
1686 bvec_iter_sectors(iter) != pick.crc.live_size ||
1688 iter.bi_sector != pos.offset));
1690 pick.ptr.offset += pick.crc.offset +
1691 (iter.bi_sector - pos.offset);
1692 pick.crc.compressed_size = bvec_iter_sectors(iter);
1693 pick.crc.uncompressed_size = bvec_iter_sectors(iter);
1694 pick.crc.offset = 0;
1695 pick.crc.live_size = bvec_iter_sectors(iter);
1696 pos.offset = iter.bi_sector;
1700 /* promote already allocated bounce rbio */
1701 } else if (bounce) {
1702 unsigned sectors = pick.crc.compressed_size;
1704 rbio = rbio_init(bio_alloc_bioset(GFP_NOIO,
1705 DIV_ROUND_UP(sectors, PAGE_SECTORS),
1706 &c->bio_read_split),
1709 bch2_bio_alloc_pages_pool(c, &rbio->bio, sectors << 9);
1710 rbio->bounce = true;
1712 } else if (flags & BCH_READ_MUST_CLONE) {
1714 * Have to clone if there were any splits, due to error
1715 * reporting issues (if a split errored, and retrying didn't
1716 * work, when it reports the error to its parent (us) we don't
1717 * know if the error was from our bio, and we should retry, or
1718 * from the whole bio, in which case we don't want to retry and
1721 rbio = rbio_init(bio_clone_fast(&orig->bio, GFP_NOIO,
1722 &c->bio_read_split),
1724 rbio->bio.bi_iter = iter;
1729 rbio->bio.bi_iter = iter;
1730 BUG_ON(bio_flagged(&rbio->bio, BIO_CHAIN));
1733 BUG_ON(bio_sectors(&rbio->bio) != pick.crc.compressed_size);
1736 rbio->submit_time = local_clock();
1738 rbio->parent = orig;
1740 rbio->end_io = orig->bio.bi_end_io;
1741 rbio->bvec_iter = iter;
1742 rbio->flags = flags;
1743 rbio->have_ioref = pick_ret > 0 && bch2_dev_get_ioref(ca, READ);
1744 rbio->narrow_crcs = narrow_crcs;
1748 rbio->devs_have = bch2_bkey_devs(k);
1751 rbio->version = k.k->version;
1752 rbio->promote = promote;
1753 INIT_WORK(&rbio->work, NULL);
1755 rbio->bio.bi_opf = orig->bio.bi_opf;
1756 rbio->bio.bi_iter.bi_sector = pick.ptr.offset;
1757 rbio->bio.bi_end_io = bch2_read_endio;
1760 trace_read_bounce(&rbio->bio);
1762 bch2_increment_clock(c, bio_sectors(&rbio->bio), READ);
1764 percpu_down_read_preempt_disable(&c->mark_lock);
1765 bucket_io_clock_reset(c, ca, PTR_BUCKET_NR(ca, &pick.ptr), READ);
1766 percpu_up_read_preempt_enable(&c->mark_lock);
1768 if (likely(!(flags & (BCH_READ_IN_RETRY|BCH_READ_LAST_FRAGMENT)))) {
1769 bio_inc_remaining(&orig->bio);
1770 trace_read_split(&orig->bio);
1773 if (!rbio->pick.idx) {
1774 if (!rbio->have_ioref) {
1775 __bcache_io_error(c, "no device to read from");
1776 bch2_rbio_error(rbio, READ_RETRY_AVOID, BLK_STS_IOERR);
1780 this_cpu_add(ca->io_done->sectors[READ][BCH_DATA_USER],
1781 bio_sectors(&rbio->bio));
1782 bio_set_dev(&rbio->bio, ca->disk_sb.bdev);
1784 if (likely(!(flags & BCH_READ_IN_RETRY)))
1785 submit_bio(&rbio->bio);
1787 submit_bio_wait(&rbio->bio);
1789 /* Attempting reconstruct read: */
1790 if (bch2_ec_read_extent(c, rbio)) {
1791 bch2_rbio_error(rbio, READ_RETRY_AVOID, BLK_STS_IOERR);
1795 if (likely(!(flags & BCH_READ_IN_RETRY)))
1796 bio_endio(&rbio->bio);
1799 if (likely(!(flags & BCH_READ_IN_RETRY))) {
1804 rbio->context = RBIO_CONTEXT_UNBOUND;
1805 bch2_read_endio(&rbio->bio);
1808 rbio = bch2_rbio_free(rbio);
1810 if (ret == READ_RETRY_AVOID) {
1811 bch2_mark_io_failure(failed, &pick);
1819 if (flags & BCH_READ_IN_RETRY)
1822 orig->bio.bi_status = BLK_STS_IOERR;
1827 * won't normally happen in the BCH_READ_NODECODE
1828 * (bch2_move_extent()) path, but if we retry and the extent we wanted
1829 * to read no longer exists we have to signal that:
1831 if (flags & BCH_READ_NODECODE)
1834 zero_fill_bio_iter(&orig->bio, iter);
1836 if (flags & BCH_READ_LAST_FRAGMENT)
1837 bch2_rbio_done(orig);
1841 void bch2_read(struct bch_fs *c, struct bch_read_bio *rbio, u64 inode)
1843 struct btree_trans trans;
1844 struct btree_iter *iter;
1846 unsigned flags = BCH_READ_RETRY_IF_STALE|
1847 BCH_READ_MAY_PROMOTE|
1848 BCH_READ_USER_MAPPED;
1851 bch2_trans_init(&trans, c);
1853 BUG_ON(rbio->_state);
1854 BUG_ON(flags & BCH_READ_NODECODE);
1855 BUG_ON(flags & BCH_READ_IN_RETRY);
1858 rbio->start_time = local_clock();
1860 for_each_btree_key(&trans, iter, BTREE_ID_EXTENTS,
1861 POS(inode, rbio->bio.bi_iter.bi_sector),
1862 BTREE_ITER_SLOTS, k, ret) {
1867 * Unlock the iterator while the btree node's lock is still in
1868 * cache, before doing the IO:
1870 bkey_reassemble(&tmp.k, k);
1871 k = bkey_i_to_s_c(&tmp.k);
1872 bch2_btree_trans_unlock(&trans);
1874 bytes = min_t(unsigned, rbio->bio.bi_iter.bi_size,
1875 (k.k->p.offset - rbio->bio.bi_iter.bi_sector) << 9);
1876 swap(rbio->bio.bi_iter.bi_size, bytes);
1878 if (rbio->bio.bi_iter.bi_size == bytes)
1879 flags |= BCH_READ_LAST_FRAGMENT;
1881 bch2_read_extent(c, rbio, k, flags);
1883 if (flags & BCH_READ_LAST_FRAGMENT)
1886 swap(rbio->bio.bi_iter.bi_size, bytes);
1887 bio_advance(&rbio->bio, bytes);
1891 * If we get here, it better have been because there was an error
1892 * reading a btree node
1895 bcache_io_error(c, &rbio->bio, "btree IO error: %i", ret);
1897 bch2_trans_exit(&trans);
1898 bch2_rbio_done(rbio);
1901 void bch2_fs_io_exit(struct bch_fs *c)
1903 if (c->promote_table.tbl)
1904 rhashtable_destroy(&c->promote_table);
1905 mempool_exit(&c->bio_bounce_pages);
1906 bioset_exit(&c->bio_write);
1907 bioset_exit(&c->bio_read_split);
1908 bioset_exit(&c->bio_read);
1911 int bch2_fs_io_init(struct bch_fs *c)
1913 if (bioset_init(&c->bio_read, 1, offsetof(struct bch_read_bio, bio),
1914 BIOSET_NEED_BVECS) ||
1915 bioset_init(&c->bio_read_split, 1, offsetof(struct bch_read_bio, bio),
1916 BIOSET_NEED_BVECS) ||
1917 bioset_init(&c->bio_write, 1, offsetof(struct bch_write_bio, bio),
1918 BIOSET_NEED_BVECS) ||
1919 mempool_init_page_pool(&c->bio_bounce_pages,
1921 c->opts.btree_node_size,
1922 c->sb.encoded_extent_max) /
1924 rhashtable_init(&c->promote_table, &bch_promote_params))