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, BTREE_ITER_MAX, 256);
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 bch_fs *c = op->c;
436 struct bkey_i_extent *e = bkey_extent_init(op->insert_keys.top);
437 struct extent_ptr_decoded p = { .crc = crc };
438 struct open_bucket *ob;
441 op->pos.offset += crc.uncompressed_size;
443 e->k.size = crc.uncompressed_size;
444 e->k.version = version;
446 BUG_ON(crc.compressed_size > wp->sectors_free);
447 wp->sectors_free -= crc.compressed_size;
449 open_bucket_for_each(c, &wp->ptrs, ob, i) {
450 struct bch_dev *ca = bch_dev_bkey_exists(c, ob->ptr.dev);
453 p.ptr.cached = !ca->mi.durability ||
454 (op->flags & BCH_WRITE_CACHED) != 0;
455 p.ptr.offset += ca->mi.bucket_size - ob->sectors_free;
456 bch2_extent_ptr_decoded_append(e, &p);
458 BUG_ON(crc.compressed_size > ob->sectors_free);
459 ob->sectors_free -= crc.compressed_size;
462 bch2_keylist_push(&op->insert_keys);
465 static struct bio *bch2_write_bio_alloc(struct bch_fs *c,
466 struct write_point *wp,
468 bool *page_alloc_failed,
471 struct bch_write_bio *wbio;
473 unsigned output_available =
474 min(wp->sectors_free << 9, src->bi_iter.bi_size);
475 unsigned pages = DIV_ROUND_UP(output_available, PAGE_SIZE);
477 bio = bio_alloc_bioset(GFP_NOIO, pages, &c->bio_write);
478 wbio = wbio_init(bio);
479 wbio->put_bio = true;
480 /* copy WRITE_SYNC flag */
481 wbio->bio.bi_opf = src->bi_opf;
484 bio->bi_iter.bi_size = output_available;
485 bch2_bio_map(bio, buf);
492 * We can't use mempool for more than c->sb.encoded_extent_max
493 * worth of pages, but we'd like to allocate more if we can:
495 while (bio->bi_iter.bi_size < output_available) {
496 unsigned len = min_t(unsigned, PAGE_SIZE,
497 output_available - bio->bi_iter.bi_size);
500 p = alloc_page(GFP_NOIO);
503 min_t(unsigned, output_available,
504 c->sb.encoded_extent_max << 9);
506 if (bio_sectors(bio) < pool_max)
507 bch2_bio_alloc_pages_pool(c, bio, pool_max);
511 bio->bi_io_vec[bio->bi_vcnt++] = (struct bio_vec) {
516 bio->bi_iter.bi_size += len;
519 *page_alloc_failed = bio->bi_vcnt < pages;
523 static int bch2_write_rechecksum(struct bch_fs *c,
524 struct bch_write_op *op,
525 unsigned new_csum_type)
527 struct bio *bio = &op->wbio.bio;
528 struct bch_extent_crc_unpacked new_crc;
531 /* bch2_rechecksum_bio() can't encrypt or decrypt data: */
533 if (bch2_csum_type_is_encryption(op->crc.csum_type) !=
534 bch2_csum_type_is_encryption(new_csum_type))
535 new_csum_type = op->crc.csum_type;
537 ret = bch2_rechecksum_bio(c, bio, op->version, op->crc,
539 op->crc.offset, op->crc.live_size,
544 bio_advance(bio, op->crc.offset << 9);
545 bio->bi_iter.bi_size = op->crc.live_size << 9;
550 static int bch2_write_decrypt(struct bch_write_op *op)
552 struct bch_fs *c = op->c;
553 struct nonce nonce = extent_nonce(op->version, op->crc);
554 struct bch_csum csum;
556 if (!bch2_csum_type_is_encryption(op->crc.csum_type))
560 * If we need to decrypt data in the write path, we'll no longer be able
561 * to verify the existing checksum (poly1305 mac, in this case) after
562 * it's decrypted - this is the last point we'll be able to reverify the
565 csum = bch2_checksum_bio(c, op->crc.csum_type, nonce, &op->wbio.bio);
566 if (bch2_crc_cmp(op->crc.csum, csum))
569 bch2_encrypt_bio(c, op->crc.csum_type, nonce, &op->wbio.bio);
570 op->crc.csum_type = 0;
571 op->crc.csum = (struct bch_csum) { 0, 0 };
575 static enum prep_encoded_ret {
578 PREP_ENCODED_CHECKSUM_ERR,
579 PREP_ENCODED_DO_WRITE,
580 } bch2_write_prep_encoded_data(struct bch_write_op *op, struct write_point *wp)
582 struct bch_fs *c = op->c;
583 struct bio *bio = &op->wbio.bio;
585 if (!(op->flags & BCH_WRITE_DATA_ENCODED))
586 return PREP_ENCODED_OK;
588 BUG_ON(bio_sectors(bio) != op->crc.compressed_size);
590 /* Can we just write the entire extent as is? */
591 if (op->crc.uncompressed_size == op->crc.live_size &&
592 op->crc.compressed_size <= wp->sectors_free &&
593 op->crc.compression_type == op->compression_type) {
594 if (!op->crc.compression_type &&
595 op->csum_type != op->crc.csum_type &&
596 bch2_write_rechecksum(c, op, op->csum_type))
597 return PREP_ENCODED_CHECKSUM_ERR;
599 return PREP_ENCODED_DO_WRITE;
603 * If the data is compressed and we couldn't write the entire extent as
604 * is, we have to decompress it:
606 if (op->crc.compression_type) {
607 struct bch_csum csum;
609 if (bch2_write_decrypt(op))
610 return PREP_ENCODED_CHECKSUM_ERR;
612 /* Last point we can still verify checksum: */
613 csum = bch2_checksum_bio(c, op->crc.csum_type,
614 extent_nonce(op->version, op->crc),
616 if (bch2_crc_cmp(op->crc.csum, csum))
617 return PREP_ENCODED_CHECKSUM_ERR;
619 if (bch2_bio_uncompress_inplace(c, bio, &op->crc))
620 return PREP_ENCODED_ERR;
624 * No longer have compressed data after this point - data might be
629 * If the data is checksummed and we're only writing a subset,
630 * rechecksum and adjust bio to point to currently live data:
632 if ((op->crc.live_size != op->crc.uncompressed_size ||
633 op->crc.csum_type != op->csum_type) &&
634 bch2_write_rechecksum(c, op, op->csum_type))
635 return PREP_ENCODED_CHECKSUM_ERR;
638 * If we want to compress the data, it has to be decrypted:
640 if ((op->compression_type ||
641 bch2_csum_type_is_encryption(op->crc.csum_type) !=
642 bch2_csum_type_is_encryption(op->csum_type)) &&
643 bch2_write_decrypt(op))
644 return PREP_ENCODED_CHECKSUM_ERR;
646 return PREP_ENCODED_OK;
649 static int bch2_write_extent(struct bch_write_op *op, struct write_point *wp)
651 struct bch_fs *c = op->c;
652 struct bio *src = &op->wbio.bio, *dst = src;
653 struct bvec_iter saved_iter;
654 struct bkey_i *key_to_write;
656 unsigned key_to_write_offset = op->insert_keys.top_p -
657 op->insert_keys.keys_p;
658 unsigned total_output = 0, total_input = 0;
660 bool page_alloc_failed = false;
663 BUG_ON(!bio_sectors(src));
665 ec_buf = bch2_writepoint_ec_buf(c, wp);
667 switch (bch2_write_prep_encoded_data(op, wp)) {
668 case PREP_ENCODED_OK:
670 case PREP_ENCODED_ERR:
673 case PREP_ENCODED_CHECKSUM_ERR:
675 case PREP_ENCODED_DO_WRITE:
677 dst = bch2_write_bio_alloc(c, wp, src,
680 bio_copy_data(dst, src);
683 init_append_extent(op, wp, op->version, op->crc);
688 op->compression_type ||
690 !(op->flags & BCH_WRITE_PAGES_STABLE)) ||
691 (bch2_csum_type_is_encryption(op->csum_type) &&
692 !(op->flags & BCH_WRITE_PAGES_OWNED))) {
693 dst = bch2_write_bio_alloc(c, wp, src,
699 saved_iter = dst->bi_iter;
702 struct bch_extent_crc_unpacked crc =
703 (struct bch_extent_crc_unpacked) { 0 };
704 struct bversion version = op->version;
705 size_t dst_len, src_len;
707 if (page_alloc_failed &&
708 bio_sectors(dst) < wp->sectors_free &&
709 bio_sectors(dst) < c->sb.encoded_extent_max)
712 BUG_ON(op->compression_type &&
713 (op->flags & BCH_WRITE_DATA_ENCODED) &&
714 bch2_csum_type_is_encryption(op->crc.csum_type));
715 BUG_ON(op->compression_type && !bounce);
717 crc.compression_type = op->compression_type
718 ? bch2_bio_compress(c, dst, &dst_len, src, &src_len,
719 op->compression_type)
721 if (!crc.compression_type) {
722 dst_len = min(dst->bi_iter.bi_size, src->bi_iter.bi_size);
723 dst_len = min_t(unsigned, dst_len, wp->sectors_free << 9);
726 dst_len = min_t(unsigned, dst_len,
727 c->sb.encoded_extent_max << 9);
730 swap(dst->bi_iter.bi_size, dst_len);
731 bio_copy_data(dst, src);
732 swap(dst->bi_iter.bi_size, dst_len);
738 BUG_ON(!src_len || !dst_len);
740 if (bch2_csum_type_is_encryption(op->csum_type)) {
741 if (bversion_zero(version)) {
742 version.lo = atomic64_inc_return(&c->key_version) + 1;
744 crc.nonce = op->nonce;
745 op->nonce += src_len >> 9;
749 if ((op->flags & BCH_WRITE_DATA_ENCODED) &&
750 !crc.compression_type &&
751 bch2_csum_type_is_encryption(op->crc.csum_type) ==
752 bch2_csum_type_is_encryption(op->csum_type)) {
754 * Note: when we're using rechecksum(), we need to be
755 * checksumming @src because it has all the data our
756 * existing checksum covers - if we bounced (because we
757 * were trying to compress), @dst will only have the
758 * part of the data the new checksum will cover.
760 * But normally we want to be checksumming post bounce,
761 * because part of the reason for bouncing is so the
762 * data can't be modified (by userspace) while it's in
765 if (bch2_rechecksum_bio(c, src, version, op->crc,
768 bio_sectors(src) - (src_len >> 9),
772 if ((op->flags & BCH_WRITE_DATA_ENCODED) &&
773 bch2_rechecksum_bio(c, src, version, op->crc,
776 bio_sectors(src) - (src_len >> 9),
780 crc.compressed_size = dst_len >> 9;
781 crc.uncompressed_size = src_len >> 9;
782 crc.live_size = src_len >> 9;
784 swap(dst->bi_iter.bi_size, dst_len);
785 bch2_encrypt_bio(c, op->csum_type,
786 extent_nonce(version, crc), dst);
787 crc.csum = bch2_checksum_bio(c, op->csum_type,
788 extent_nonce(version, crc), dst);
789 crc.csum_type = op->csum_type;
790 swap(dst->bi_iter.bi_size, dst_len);
793 init_append_extent(op, wp, version, crc);
796 bio_advance(dst, dst_len);
797 bio_advance(src, src_len);
798 total_output += dst_len;
799 total_input += src_len;
800 } while (dst->bi_iter.bi_size &&
801 src->bi_iter.bi_size &&
803 !bch2_keylist_realloc(&op->insert_keys,
805 ARRAY_SIZE(op->inline_keys),
806 BKEY_EXTENT_U64s_MAX));
808 more = src->bi_iter.bi_size != 0;
810 dst->bi_iter = saved_iter;
812 if (dst == src && more) {
813 BUG_ON(total_output != total_input);
815 dst = bio_split(src, total_input >> 9,
816 GFP_NOIO, &c->bio_write);
817 wbio_init(dst)->put_bio = true;
818 /* copy WRITE_SYNC flag */
819 dst->bi_opf = src->bi_opf;
822 dst->bi_iter.bi_size = total_output;
824 /* Free unneeded pages after compressing: */
825 if (to_wbio(dst)->bounce)
826 while (dst->bi_vcnt > DIV_ROUND_UP(dst->bi_iter.bi_size, PAGE_SIZE))
827 mempool_free(dst->bi_io_vec[--dst->bi_vcnt].bv_page,
828 &c->bio_bounce_pages);
830 /* might have done a realloc... */
832 key_to_write = (void *) (op->insert_keys.keys_p + key_to_write_offset);
834 bch2_ec_add_backpointer(c, wp,
835 bkey_start_pos(&key_to_write->k),
838 dst->bi_end_io = bch2_write_endio;
839 dst->bi_private = &op->cl;
840 bio_set_op_attrs(dst, REQ_OP_WRITE, 0);
842 closure_get(dst->bi_private);
844 bch2_submit_wbio_replicas(to_wbio(dst), c, BCH_DATA_USER,
848 bch_err(c, "error verifying existing checksum while "
849 "rewriting existing data (memory corruption?)");
852 if (to_wbio(dst)->bounce)
853 bch2_bio_free_pages_pool(c, dst);
854 if (to_wbio(dst)->put_bio)
860 static void __bch2_write(struct closure *cl)
862 struct bch_write_op *op = container_of(cl, struct bch_write_op, cl);
863 struct bch_fs *c = op->c;
864 struct write_point *wp;
867 memset(&op->failed, 0, sizeof(op->failed));
870 /* +1 for possible cache device: */
871 if (op->open_buckets.nr + op->nr_replicas + 1 >
872 ARRAY_SIZE(op->open_buckets.v))
875 if (bch2_keylist_realloc(&op->insert_keys,
877 ARRAY_SIZE(op->inline_keys),
878 BKEY_EXTENT_U64s_MAX))
881 wp = bch2_alloc_sectors_start(c,
883 op->opts.erasure_code,
887 op->nr_replicas_required,
890 (op->flags & BCH_WRITE_ALLOC_NOWAIT) ? NULL : cl);
893 if (unlikely(IS_ERR(wp))) {
894 if (unlikely(PTR_ERR(wp) != -EAGAIN)) {
902 ret = bch2_write_extent(op, wp);
904 bch2_open_bucket_get(c, wp, &op->open_buckets);
905 bch2_alloc_sectors_done(c, wp);
911 continue_at(cl, bch2_write_index, index_update_wq(op));
916 continue_at(cl, !bch2_keylist_empty(&op->insert_keys)
918 : bch2_write_done, index_update_wq(op));
923 if (!bch2_keylist_empty(&op->insert_keys)) {
924 __bch2_write_index(op);
927 continue_at_nobarrier(cl, bch2_write_done, NULL);
936 * bch_write - handle a write to a cache device or flash only volume
938 * This is the starting point for any data to end up in a cache device; it could
939 * be from a normal write, or a writeback write, or a write to a flash only
940 * volume - it's also used by the moving garbage collector to compact data in
941 * mostly empty buckets.
943 * It first writes the data to the cache, creating a list of keys to be inserted
944 * (if the data won't fit in a single open bucket, there will be multiple keys);
945 * after the data is written it calls bch_journal, and after the keys have been
946 * added to the next journal write they're inserted into the btree.
948 * If op->discard is true, instead of inserting the data it invalidates the
949 * region of the cache represented by op->bio and op->inode.
951 void bch2_write(struct closure *cl)
953 struct bch_write_op *op = container_of(cl, struct bch_write_op, cl);
954 struct bch_fs *c = op->c;
956 BUG_ON(!op->nr_replicas);
957 BUG_ON(!op->write_point.v);
958 BUG_ON(!bkey_cmp(op->pos, POS_MAX));
959 BUG_ON(bio_sectors(&op->wbio.bio) > U16_MAX);
961 op->start_time = local_clock();
963 bch2_keylist_init(&op->insert_keys, op->inline_keys);
964 wbio_init(&op->wbio.bio)->put_bio = false;
966 if (c->opts.nochanges ||
967 !percpu_ref_tryget(&c->writes)) {
968 __bcache_io_error(c, "read only");
970 if (!(op->flags & BCH_WRITE_NOPUT_RESERVATION))
971 bch2_disk_reservation_put(c, &op->res);
976 bch2_increment_clock(c, bio_sectors(&op->wbio.bio), WRITE);
978 continue_at_nobarrier(cl, __bch2_write, NULL);
981 /* Cache promotion on read */
988 struct rhash_head hash;
991 struct migrate_write write;
992 struct bio_vec bi_inline_vecs[0]; /* must be last */
995 static const struct rhashtable_params bch_promote_params = {
996 .head_offset = offsetof(struct promote_op, hash),
997 .key_offset = offsetof(struct promote_op, pos),
998 .key_len = sizeof(struct bpos),
1001 static inline bool should_promote(struct bch_fs *c, struct bkey_s_c k,
1003 struct bch_io_opts opts,
1006 if (!opts.promote_target)
1009 if (!(flags & BCH_READ_MAY_PROMOTE))
1012 if (percpu_ref_is_dying(&c->writes))
1015 if (!bkey_extent_is_data(k.k))
1018 if (bch2_extent_has_target(c, bkey_s_c_to_extent(k), opts.promote_target))
1021 if (bch2_target_congested(c, opts.promote_target))
1024 if (rhashtable_lookup_fast(&c->promote_table, &pos,
1025 bch_promote_params))
1031 static void promote_free(struct bch_fs *c, struct promote_op *op)
1035 ret = rhashtable_remove_fast(&c->promote_table, &op->hash,
1036 bch_promote_params);
1038 percpu_ref_put(&c->writes);
1042 static void promote_done(struct closure *cl)
1044 struct promote_op *op =
1045 container_of(cl, struct promote_op, cl);
1046 struct bch_fs *c = op->write.op.c;
1048 bch2_time_stats_update(&c->times[BCH_TIME_data_promote],
1051 bch2_bio_free_pages_pool(c, &op->write.op.wbio.bio);
1052 promote_free(c, op);
1055 static void promote_start(struct promote_op *op, struct bch_read_bio *rbio)
1057 struct bch_fs *c = rbio->c;
1058 struct closure *cl = &op->cl;
1059 struct bio *bio = &op->write.op.wbio.bio;
1061 trace_promote(&rbio->bio);
1063 /* we now own pages: */
1064 BUG_ON(!rbio->bounce);
1065 BUG_ON(rbio->bio.bi_vcnt > bio->bi_max_vecs);
1067 memcpy(bio->bi_io_vec, rbio->bio.bi_io_vec,
1068 sizeof(struct bio_vec) * rbio->bio.bi_vcnt);
1069 swap(bio->bi_vcnt, rbio->bio.bi_vcnt);
1071 bch2_migrate_read_done(&op->write, rbio);
1073 closure_init(cl, NULL);
1074 closure_call(&op->write.op.cl, bch2_write, c->wq, cl);
1075 closure_return_with_destructor(cl, promote_done);
1079 static struct promote_op *__promote_alloc(struct bch_fs *c,
1081 struct extent_ptr_decoded *pick,
1082 struct bch_io_opts opts,
1083 unsigned rbio_sectors,
1084 struct bch_read_bio **rbio)
1086 struct promote_op *op = NULL;
1088 unsigned rbio_pages = DIV_ROUND_UP(rbio_sectors, PAGE_SECTORS);
1089 /* data might have to be decompressed in the write path: */
1090 unsigned wbio_pages = DIV_ROUND_UP(pick->crc.uncompressed_size,
1094 if (!percpu_ref_tryget(&c->writes))
1097 op = kzalloc(sizeof(*op) + sizeof(struct bio_vec) * wbio_pages,
1102 op->start_time = local_clock();
1106 * promotes require bouncing, but if the extent isn't
1107 * checksummed/compressed it might be too big for the mempool:
1109 if (rbio_sectors > c->sb.encoded_extent_max) {
1110 *rbio = kzalloc(sizeof(struct bch_read_bio) +
1111 sizeof(struct bio_vec) * rbio_pages,
1116 rbio_init(&(*rbio)->bio, opts);
1117 bio_init(&(*rbio)->bio, (*rbio)->bio.bi_inline_vecs,
1120 (*rbio)->bio.bi_iter.bi_size = rbio_sectors << 9;
1121 bch2_bio_map(&(*rbio)->bio, NULL);
1123 if (bch2_bio_alloc_pages(&(*rbio)->bio, GFP_NOIO))
1126 (*rbio)->bounce = true;
1127 (*rbio)->split = true;
1128 (*rbio)->kmalloc = true;
1131 if (rhashtable_lookup_insert_fast(&c->promote_table, &op->hash,
1132 bch_promote_params))
1135 bio = &op->write.op.wbio.bio;
1136 bio_init(bio, bio->bi_inline_vecs, wbio_pages);
1138 ret = bch2_migrate_write_init(c, &op->write,
1139 writepoint_hashed((unsigned long) current),
1142 (struct data_opts) {
1143 .target = opts.promote_target
1151 bio_free_pages(&(*rbio)->bio);
1155 percpu_ref_put(&c->writes);
1159 static inline struct promote_op *promote_alloc(struct bch_fs *c,
1160 struct bvec_iter iter,
1162 struct extent_ptr_decoded *pick,
1163 struct bch_io_opts opts,
1165 struct bch_read_bio **rbio,
1169 bool promote_full = *read_full || READ_ONCE(c->promote_whole_extents);
1170 unsigned sectors = promote_full
1171 ? pick->crc.compressed_size
1172 : bvec_iter_sectors(iter);
1173 struct bpos pos = promote_full
1174 ? bkey_start_pos(k.k)
1175 : POS(k.k->p.inode, iter.bi_sector);
1176 struct promote_op *promote;
1178 if (!should_promote(c, k, pos, opts, flags))
1181 promote = __promote_alloc(c, pos, pick, opts, sectors, rbio);
1186 *read_full = promote_full;
1192 #define READ_RETRY_AVOID 1
1193 #define READ_RETRY 2
1198 RBIO_CONTEXT_HIGHPRI,
1199 RBIO_CONTEXT_UNBOUND,
1202 static inline struct bch_read_bio *
1203 bch2_rbio_parent(struct bch_read_bio *rbio)
1205 return rbio->split ? rbio->parent : rbio;
1209 static void bch2_rbio_punt(struct bch_read_bio *rbio, work_func_t fn,
1210 enum rbio_context context,
1211 struct workqueue_struct *wq)
1213 if (context <= rbio->context) {
1216 rbio->work.func = fn;
1217 rbio->context = context;
1218 queue_work(wq, &rbio->work);
1222 static inline struct bch_read_bio *bch2_rbio_free(struct bch_read_bio *rbio)
1224 BUG_ON(rbio->bounce && !rbio->split);
1227 promote_free(rbio->c, rbio->promote);
1228 rbio->promote = NULL;
1231 bch2_bio_free_pages_pool(rbio->c, &rbio->bio);
1234 struct bch_read_bio *parent = rbio->parent;
1239 bio_put(&rbio->bio);
1247 static void bch2_rbio_done(struct bch_read_bio *rbio)
1249 bch2_time_stats_update(&rbio->c->times[BCH_TIME_data_read],
1251 bio_endio(&rbio->bio);
1254 static void bch2_read_retry_nodecode(struct bch_fs *c, struct bch_read_bio *rbio,
1255 struct bvec_iter bvec_iter, u64 inode,
1256 struct bch_io_failures *failed,
1259 struct btree_trans trans;
1260 struct btree_iter *iter;
1265 flags &= ~BCH_READ_LAST_FRAGMENT;
1267 bch2_trans_init(&trans, c, 0, 0);
1269 iter = bch2_trans_get_iter(&trans, BTREE_ID_EXTENTS,
1270 rbio->pos, BTREE_ITER_SLOTS);
1272 rbio->bio.bi_status = 0;
1274 k = bch2_btree_iter_peek_slot(iter);
1278 bkey_reassemble(&tmp.k, k);
1279 k = bkey_i_to_s_c(&tmp.k);
1280 bch2_trans_unlock(&trans);
1282 if (!bkey_extent_is_data(k.k) ||
1283 !bch2_extent_matches_ptr(c, bkey_i_to_s_c_extent(&tmp.k),
1286 rbio->pick.crc.offset)) {
1287 /* extent we wanted to read no longer exists: */
1292 ret = __bch2_read_extent(c, rbio, bvec_iter, k, failed, flags);
1293 if (ret == READ_RETRY)
1298 bch2_rbio_done(rbio);
1299 bch2_trans_exit(&trans);
1302 rbio->bio.bi_status = BLK_STS_IOERR;
1306 static void bch2_read_retry(struct bch_fs *c, struct bch_read_bio *rbio,
1307 struct bvec_iter bvec_iter, u64 inode,
1308 struct bch_io_failures *failed, unsigned flags)
1310 struct btree_trans trans;
1311 struct btree_iter *iter;
1315 bch2_trans_init(&trans, c, 0, 0);
1317 flags &= ~BCH_READ_LAST_FRAGMENT;
1318 flags |= BCH_READ_MUST_CLONE;
1320 for_each_btree_key(&trans, iter, BTREE_ID_EXTENTS,
1321 POS(inode, bvec_iter.bi_sector),
1322 BTREE_ITER_SLOTS, k, ret) {
1326 bkey_reassemble(&tmp.k, k);
1327 k = bkey_i_to_s_c(&tmp.k);
1328 bch2_trans_unlock(&trans);
1330 bytes = min_t(unsigned, bvec_iter.bi_size,
1331 (k.k->p.offset - bvec_iter.bi_sector) << 9);
1332 swap(bvec_iter.bi_size, bytes);
1334 ret = __bch2_read_extent(c, rbio, bvec_iter, k, failed, flags);
1342 if (bytes == bvec_iter.bi_size)
1345 swap(bvec_iter.bi_size, bytes);
1346 bio_advance_iter(&rbio->bio, &bvec_iter, bytes);
1350 * If we get here, it better have been because there was an error
1351 * reading a btree node
1354 __bcache_io_error(c, "btree IO error: %i", ret);
1356 rbio->bio.bi_status = BLK_STS_IOERR;
1358 bch2_trans_exit(&trans);
1359 bch2_rbio_done(rbio);
1362 static void bch2_rbio_retry(struct work_struct *work)
1364 struct bch_read_bio *rbio =
1365 container_of(work, struct bch_read_bio, work);
1366 struct bch_fs *c = rbio->c;
1367 struct bvec_iter iter = rbio->bvec_iter;
1368 unsigned flags = rbio->flags;
1369 u64 inode = rbio->pos.inode;
1370 struct bch_io_failures failed = { .nr = 0 };
1372 trace_read_retry(&rbio->bio);
1374 if (rbio->retry == READ_RETRY_AVOID)
1375 bch2_mark_io_failure(&failed, &rbio->pick);
1377 rbio->bio.bi_status = 0;
1379 rbio = bch2_rbio_free(rbio);
1381 flags |= BCH_READ_IN_RETRY;
1382 flags &= ~BCH_READ_MAY_PROMOTE;
1384 if (flags & BCH_READ_NODECODE)
1385 bch2_read_retry_nodecode(c, rbio, iter, inode, &failed, flags);
1387 bch2_read_retry(c, rbio, iter, inode, &failed, flags);
1390 static void bch2_rbio_error(struct bch_read_bio *rbio, int retry,
1393 rbio->retry = retry;
1395 if (rbio->flags & BCH_READ_IN_RETRY)
1398 if (retry == READ_ERR) {
1399 rbio = bch2_rbio_free(rbio);
1401 rbio->bio.bi_status = error;
1402 bch2_rbio_done(rbio);
1404 bch2_rbio_punt(rbio, bch2_rbio_retry,
1405 RBIO_CONTEXT_UNBOUND, system_unbound_wq);
1409 static void bch2_rbio_narrow_crcs(struct bch_read_bio *rbio)
1411 struct bch_fs *c = rbio->c;
1412 struct btree_trans trans;
1413 struct btree_iter *iter;
1415 struct bkey_i_extent *e;
1417 struct bch_extent_crc_unpacked new_crc;
1418 u64 data_offset = rbio->pos.offset - rbio->pick.crc.offset;
1421 if (rbio->pick.crc.compression_type)
1424 bch2_trans_init(&trans, c, 0, 0);
1426 bch2_trans_begin(&trans);
1428 iter = bch2_trans_get_iter(&trans, BTREE_ID_EXTENTS, rbio->pos,
1430 k = bch2_btree_iter_peek(iter);
1431 if (IS_ERR_OR_NULL(k.k))
1434 if (!bkey_extent_is_data(k.k))
1437 bkey_reassemble(&new.k, k);
1438 e = bkey_i_to_extent(&new.k);
1440 if (!bch2_extent_matches_ptr(c, extent_i_to_s_c(e),
1441 rbio->pick.ptr, data_offset) ||
1442 bversion_cmp(e->k.version, rbio->version))
1445 /* Extent was merged? */
1446 if (bkey_start_offset(&e->k) < data_offset ||
1447 e->k.p.offset > data_offset + rbio->pick.crc.uncompressed_size)
1450 if (bch2_rechecksum_bio(c, &rbio->bio, rbio->version,
1451 rbio->pick.crc, NULL, &new_crc,
1452 bkey_start_offset(&e->k) - data_offset, e->k.size,
1453 rbio->pick.crc.csum_type)) {
1454 bch_err(c, "error verifying existing checksum while narrowing checksum (memory corruption?)");
1458 if (!bch2_extent_narrow_crcs(e, new_crc))
1461 bch2_trans_update(&trans, BTREE_INSERT_ENTRY(iter, &e->k_i));
1462 ret = bch2_trans_commit(&trans, NULL, NULL,
1463 BTREE_INSERT_ATOMIC|
1464 BTREE_INSERT_NOFAIL|
1465 BTREE_INSERT_NOWAIT);
1469 bch2_trans_exit(&trans);
1472 static bool should_narrow_crcs(struct bkey_s_c k,
1473 struct extent_ptr_decoded *pick,
1476 return !(flags & BCH_READ_IN_RETRY) &&
1477 bkey_extent_is_data(k.k) &&
1478 bch2_can_narrow_extent_crcs(bkey_s_c_to_extent(k), pick->crc);
1481 /* Inner part that may run in process context */
1482 static void __bch2_read_endio(struct work_struct *work)
1484 struct bch_read_bio *rbio =
1485 container_of(work, struct bch_read_bio, work);
1486 struct bch_fs *c = rbio->c;
1487 struct bch_dev *ca = bch_dev_bkey_exists(c, rbio->pick.ptr.dev);
1488 struct bio *src = &rbio->bio;
1489 struct bio *dst = &bch2_rbio_parent(rbio)->bio;
1490 struct bvec_iter dst_iter = rbio->bvec_iter;
1491 struct bch_extent_crc_unpacked crc = rbio->pick.crc;
1492 struct nonce nonce = extent_nonce(rbio->version, crc);
1493 struct bch_csum csum;
1495 /* Reset iterator for checksumming and copying bounced data: */
1497 src->bi_iter.bi_size = crc.compressed_size << 9;
1498 src->bi_iter.bi_idx = 0;
1499 src->bi_iter.bi_bvec_done = 0;
1501 src->bi_iter = rbio->bvec_iter;
1504 csum = bch2_checksum_bio(c, crc.csum_type, nonce, src);
1505 if (bch2_crc_cmp(csum, rbio->pick.crc.csum))
1508 if (unlikely(rbio->narrow_crcs))
1509 bch2_rbio_narrow_crcs(rbio);
1511 if (rbio->flags & BCH_READ_NODECODE)
1514 /* Adjust crc to point to subset of data we want: */
1515 crc.offset += rbio->bvec_iter.bi_sector - rbio->pos.offset;
1516 crc.live_size = bvec_iter_sectors(rbio->bvec_iter);
1518 if (crc.compression_type != BCH_COMPRESSION_NONE) {
1519 bch2_encrypt_bio(c, crc.csum_type, nonce, src);
1520 if (bch2_bio_uncompress(c, src, dst, dst_iter, crc))
1521 goto decompression_err;
1523 /* don't need to decrypt the entire bio: */
1524 nonce = nonce_add(nonce, crc.offset << 9);
1525 bio_advance(src, crc.offset << 9);
1527 BUG_ON(src->bi_iter.bi_size < dst_iter.bi_size);
1528 src->bi_iter.bi_size = dst_iter.bi_size;
1530 bch2_encrypt_bio(c, crc.csum_type, nonce, src);
1533 struct bvec_iter src_iter = src->bi_iter;
1534 bio_copy_data_iter(dst, &dst_iter, src, &src_iter);
1538 if (rbio->promote) {
1540 * Re encrypt data we decrypted, so it's consistent with
1543 bch2_encrypt_bio(c, crc.csum_type, nonce, src);
1544 promote_start(rbio->promote, rbio);
1545 rbio->promote = NULL;
1548 if (likely(!(rbio->flags & BCH_READ_IN_RETRY))) {
1549 rbio = bch2_rbio_free(rbio);
1550 bch2_rbio_done(rbio);
1555 * Checksum error: if the bio wasn't bounced, we may have been
1556 * reading into buffers owned by userspace (that userspace can
1557 * scribble over) - retry the read, bouncing it this time:
1559 if (!rbio->bounce && (rbio->flags & BCH_READ_USER_MAPPED)) {
1560 rbio->flags |= BCH_READ_MUST_BOUNCE;
1561 bch2_rbio_error(rbio, READ_RETRY, BLK_STS_IOERR);
1565 bch2_dev_io_error(ca,
1566 "data checksum error, inode %llu offset %llu: expected %0llx:%0llx got %0llx:%0llx (type %u)",
1567 rbio->pos.inode, (u64) rbio->bvec_iter.bi_sector,
1568 rbio->pick.crc.csum.hi, rbio->pick.crc.csum.lo,
1569 csum.hi, csum.lo, crc.csum_type);
1570 bch2_rbio_error(rbio, READ_RETRY_AVOID, BLK_STS_IOERR);
1573 __bcache_io_error(c, "decompression error, inode %llu offset %llu",
1575 (u64) rbio->bvec_iter.bi_sector);
1576 bch2_rbio_error(rbio, READ_ERR, BLK_STS_IOERR);
1580 static void bch2_read_endio(struct bio *bio)
1582 struct bch_read_bio *rbio =
1583 container_of(bio, struct bch_read_bio, bio);
1584 struct bch_fs *c = rbio->c;
1585 struct bch_dev *ca = bch_dev_bkey_exists(c, rbio->pick.ptr.dev);
1586 struct workqueue_struct *wq = NULL;
1587 enum rbio_context context = RBIO_CONTEXT_NULL;
1589 if (rbio->have_ioref) {
1590 bch2_latency_acct(ca, rbio->submit_time, READ);
1591 percpu_ref_put(&ca->io_ref);
1595 rbio->bio.bi_end_io = rbio->end_io;
1597 if (bch2_dev_io_err_on(bio->bi_status, ca, "data read")) {
1598 bch2_rbio_error(rbio, READ_RETRY_AVOID, bio->bi_status);
1602 if (rbio->pick.ptr.cached &&
1603 (((rbio->flags & BCH_READ_RETRY_IF_STALE) && race_fault()) ||
1604 ptr_stale(ca, &rbio->pick.ptr))) {
1605 atomic_long_inc(&c->read_realloc_races);
1607 if (rbio->flags & BCH_READ_RETRY_IF_STALE)
1608 bch2_rbio_error(rbio, READ_RETRY, BLK_STS_AGAIN);
1610 bch2_rbio_error(rbio, READ_ERR, BLK_STS_AGAIN);
1614 if (rbio->narrow_crcs ||
1615 rbio->pick.crc.compression_type ||
1616 bch2_csum_type_is_encryption(rbio->pick.crc.csum_type))
1617 context = RBIO_CONTEXT_UNBOUND, wq = system_unbound_wq;
1618 else if (rbio->pick.crc.csum_type)
1619 context = RBIO_CONTEXT_HIGHPRI, wq = system_highpri_wq;
1621 bch2_rbio_punt(rbio, __bch2_read_endio, context, wq);
1624 int __bch2_read_extent(struct bch_fs *c, struct bch_read_bio *orig,
1625 struct bvec_iter iter, struct bkey_s_c k,
1626 struct bch_io_failures *failed, unsigned flags)
1628 struct extent_ptr_decoded pick;
1629 struct bch_read_bio *rbio = NULL;
1631 struct promote_op *promote = NULL;
1632 bool bounce = false, read_full = false, narrow_crcs = false;
1633 struct bpos pos = bkey_start_pos(k.k);
1636 pick_ret = bch2_bkey_pick_read_device(c, k, failed, &pick);
1638 /* hole or reservation - just zero fill: */
1643 __bcache_io_error(c, "no device to read from");
1648 ca = bch_dev_bkey_exists(c, pick.ptr.dev);
1650 if (flags & BCH_READ_NODECODE) {
1652 * can happen if we retry, and the extent we were going to read
1653 * has been merged in the meantime:
1655 if (pick.crc.compressed_size > orig->bio.bi_vcnt * PAGE_SECTORS)
1658 iter.bi_sector = pos.offset;
1659 iter.bi_size = pick.crc.compressed_size << 9;
1663 if (!(flags & BCH_READ_LAST_FRAGMENT) ||
1664 bio_flagged(&orig->bio, BIO_CHAIN))
1665 flags |= BCH_READ_MUST_CLONE;
1667 narrow_crcs = should_narrow_crcs(k, &pick, flags);
1669 if (narrow_crcs && (flags & BCH_READ_USER_MAPPED))
1670 flags |= BCH_READ_MUST_BOUNCE;
1672 EBUG_ON(bkey_start_offset(k.k) > iter.bi_sector ||
1673 k.k->p.offset < bvec_iter_end_sector(iter));
1675 if (pick.crc.compression_type != BCH_COMPRESSION_NONE ||
1676 (pick.crc.csum_type != BCH_CSUM_NONE &&
1677 (bvec_iter_sectors(iter) != pick.crc.uncompressed_size ||
1678 (bch2_csum_type_is_encryption(pick.crc.csum_type) &&
1679 (flags & BCH_READ_USER_MAPPED)) ||
1680 (flags & BCH_READ_MUST_BOUNCE)))) {
1685 promote = promote_alloc(c, iter, k, &pick, orig->opts, flags,
1686 &rbio, &bounce, &read_full);
1689 EBUG_ON(pick.crc.compression_type);
1690 EBUG_ON(pick.crc.csum_type &&
1691 (bvec_iter_sectors(iter) != pick.crc.uncompressed_size ||
1692 bvec_iter_sectors(iter) != pick.crc.live_size ||
1694 iter.bi_sector != pos.offset));
1696 pick.ptr.offset += pick.crc.offset +
1697 (iter.bi_sector - pos.offset);
1698 pick.crc.compressed_size = bvec_iter_sectors(iter);
1699 pick.crc.uncompressed_size = bvec_iter_sectors(iter);
1700 pick.crc.offset = 0;
1701 pick.crc.live_size = bvec_iter_sectors(iter);
1702 pos.offset = iter.bi_sector;
1706 /* promote already allocated bounce rbio */
1707 } else if (bounce) {
1708 unsigned sectors = pick.crc.compressed_size;
1710 rbio = rbio_init(bio_alloc_bioset(GFP_NOIO,
1711 DIV_ROUND_UP(sectors, PAGE_SECTORS),
1712 &c->bio_read_split),
1715 bch2_bio_alloc_pages_pool(c, &rbio->bio, sectors << 9);
1716 rbio->bounce = true;
1718 } else if (flags & BCH_READ_MUST_CLONE) {
1720 * Have to clone if there were any splits, due to error
1721 * reporting issues (if a split errored, and retrying didn't
1722 * work, when it reports the error to its parent (us) we don't
1723 * know if the error was from our bio, and we should retry, or
1724 * from the whole bio, in which case we don't want to retry and
1727 rbio = rbio_init(bio_clone_fast(&orig->bio, GFP_NOIO,
1728 &c->bio_read_split),
1730 rbio->bio.bi_iter = iter;
1735 rbio->bio.bi_iter = iter;
1736 BUG_ON(bio_flagged(&rbio->bio, BIO_CHAIN));
1739 BUG_ON(bio_sectors(&rbio->bio) != pick.crc.compressed_size);
1742 rbio->submit_time = local_clock();
1744 rbio->parent = orig;
1746 rbio->end_io = orig->bio.bi_end_io;
1747 rbio->bvec_iter = iter;
1748 rbio->flags = flags;
1749 rbio->have_ioref = pick_ret > 0 && bch2_dev_get_ioref(ca, READ);
1750 rbio->narrow_crcs = narrow_crcs;
1754 rbio->devs_have = bch2_bkey_devs(k);
1757 rbio->version = k.k->version;
1758 rbio->promote = promote;
1759 INIT_WORK(&rbio->work, NULL);
1761 rbio->bio.bi_opf = orig->bio.bi_opf;
1762 rbio->bio.bi_iter.bi_sector = pick.ptr.offset;
1763 rbio->bio.bi_end_io = bch2_read_endio;
1766 trace_read_bounce(&rbio->bio);
1768 bch2_increment_clock(c, bio_sectors(&rbio->bio), READ);
1770 percpu_down_read_preempt_disable(&c->mark_lock);
1771 bucket_io_clock_reset(c, ca, PTR_BUCKET_NR(ca, &pick.ptr), READ);
1772 percpu_up_read_preempt_enable(&c->mark_lock);
1774 if (likely(!(flags & (BCH_READ_IN_RETRY|BCH_READ_LAST_FRAGMENT)))) {
1775 bio_inc_remaining(&orig->bio);
1776 trace_read_split(&orig->bio);
1779 if (!rbio->pick.idx) {
1780 if (!rbio->have_ioref) {
1781 __bcache_io_error(c, "no device to read from");
1782 bch2_rbio_error(rbio, READ_RETRY_AVOID, BLK_STS_IOERR);
1786 this_cpu_add(ca->io_done->sectors[READ][BCH_DATA_USER],
1787 bio_sectors(&rbio->bio));
1788 bio_set_dev(&rbio->bio, ca->disk_sb.bdev);
1790 if (likely(!(flags & BCH_READ_IN_RETRY)))
1791 submit_bio(&rbio->bio);
1793 submit_bio_wait(&rbio->bio);
1795 /* Attempting reconstruct read: */
1796 if (bch2_ec_read_extent(c, rbio)) {
1797 bch2_rbio_error(rbio, READ_RETRY_AVOID, BLK_STS_IOERR);
1801 if (likely(!(flags & BCH_READ_IN_RETRY)))
1802 bio_endio(&rbio->bio);
1805 if (likely(!(flags & BCH_READ_IN_RETRY))) {
1810 rbio->context = RBIO_CONTEXT_UNBOUND;
1811 bch2_read_endio(&rbio->bio);
1814 rbio = bch2_rbio_free(rbio);
1816 if (ret == READ_RETRY_AVOID) {
1817 bch2_mark_io_failure(failed, &pick);
1825 if (flags & BCH_READ_IN_RETRY)
1828 orig->bio.bi_status = BLK_STS_IOERR;
1833 * won't normally happen in the BCH_READ_NODECODE
1834 * (bch2_move_extent()) path, but if we retry and the extent we wanted
1835 * to read no longer exists we have to signal that:
1837 if (flags & BCH_READ_NODECODE)
1840 zero_fill_bio_iter(&orig->bio, iter);
1842 if (flags & BCH_READ_LAST_FRAGMENT)
1843 bch2_rbio_done(orig);
1847 void bch2_read(struct bch_fs *c, struct bch_read_bio *rbio, u64 inode)
1849 struct btree_trans trans;
1850 struct btree_iter *iter;
1852 unsigned flags = BCH_READ_RETRY_IF_STALE|
1853 BCH_READ_MAY_PROMOTE|
1854 BCH_READ_USER_MAPPED;
1857 bch2_trans_init(&trans, c, 0, 0);
1859 BUG_ON(rbio->_state);
1860 BUG_ON(flags & BCH_READ_NODECODE);
1861 BUG_ON(flags & BCH_READ_IN_RETRY);
1864 rbio->start_time = local_clock();
1866 for_each_btree_key(&trans, iter, BTREE_ID_EXTENTS,
1867 POS(inode, rbio->bio.bi_iter.bi_sector),
1868 BTREE_ITER_SLOTS, k, ret) {
1873 * Unlock the iterator while the btree node's lock is still in
1874 * cache, before doing the IO:
1876 bkey_reassemble(&tmp.k, k);
1877 k = bkey_i_to_s_c(&tmp.k);
1878 bch2_trans_unlock(&trans);
1880 bytes = min_t(unsigned, rbio->bio.bi_iter.bi_size,
1881 (k.k->p.offset - rbio->bio.bi_iter.bi_sector) << 9);
1882 swap(rbio->bio.bi_iter.bi_size, bytes);
1884 if (rbio->bio.bi_iter.bi_size == bytes)
1885 flags |= BCH_READ_LAST_FRAGMENT;
1887 bch2_read_extent(c, rbio, k, flags);
1889 if (flags & BCH_READ_LAST_FRAGMENT)
1892 swap(rbio->bio.bi_iter.bi_size, bytes);
1893 bio_advance(&rbio->bio, bytes);
1897 * If we get here, it better have been because there was an error
1898 * reading a btree node
1901 bcache_io_error(c, &rbio->bio, "btree IO error: %i", ret);
1903 bch2_trans_exit(&trans);
1904 bch2_rbio_done(rbio);
1907 void bch2_fs_io_exit(struct bch_fs *c)
1909 if (c->promote_table.tbl)
1910 rhashtable_destroy(&c->promote_table);
1911 mempool_exit(&c->bio_bounce_pages);
1912 bioset_exit(&c->bio_write);
1913 bioset_exit(&c->bio_read_split);
1914 bioset_exit(&c->bio_read);
1917 int bch2_fs_io_init(struct bch_fs *c)
1919 if (bioset_init(&c->bio_read, 1, offsetof(struct bch_read_bio, bio),
1920 BIOSET_NEED_BVECS) ||
1921 bioset_init(&c->bio_read_split, 1, offsetof(struct bch_read_bio, bio),
1922 BIOSET_NEED_BVECS) ||
1923 bioset_init(&c->bio_write, 1, offsetof(struct bch_write_bio, bio),
1924 BIOSET_NEED_BVECS) ||
1925 mempool_init_page_pool(&c->bio_bounce_pages,
1927 c->opts.btree_node_size,
1928 c->sb.encoded_extent_max) /
1930 rhashtable_init(&c->promote_table, &bch_promote_params))