X-Git-Url: https://git.sesse.net/?a=blobdiff_plain;f=libbcachefs%2Fio.c;h=1ad4c7d77812998a17d4fec011ce645176d49d81;hb=6f5afc0c12bbf56ffdabe5b2c5297aef255c4baa;hp=021a80df098fc29cc44f24c218f77469b8c888d4;hpb=55fbb25501330038e1714905b9ddeb25d875c11c;p=bcachefs-tools-debian diff --git a/libbcachefs/io.c b/libbcachefs/io.c index 021a80d..1ad4c7d 100644 --- a/libbcachefs/io.c +++ b/libbcachefs/io.c @@ -1,3 +1,4 @@ +// SPDX-License-Identifier: GPL-2.0 /* * Some low level IO code, and hacks for various block layer limitations * @@ -6,7 +7,9 @@ */ #include "bcachefs.h" +#include "alloc_background.h" #include "alloc_foreground.h" +#include "bkey_buf.h" #include "bset.h" #include "btree_update.h" #include "buckets.h" @@ -15,22 +18,32 @@ #include "clock.h" #include "debug.h" #include "disk_groups.h" +#include "ec.h" #include "error.h" -#include "extents.h" +#include "extent_update.h" +#include "inode.h" #include "io.h" #include "journal.h" #include "keylist.h" #include "move.h" #include "rebalance.h" -#include "replicas.h" +#include "subvolume.h" #include "super.h" #include "super-io.h" #include #include +#include #include +const char *bch2_blk_status_to_str(blk_status_t status) +{ + if (status == BLK_STS_REMOVED) + return "device removed"; + return blk_status_to_str(status); +} + static bool bch2_target_congested(struct bch_fs *c, u16 target) { const struct bch_devs_mask *devs; @@ -43,7 +56,9 @@ static bool bch2_target_congested(struct bch_fs *c, u16 target) return false; rcu_read_lock(); - devs = bch2_target_to_mask(c, target); + devs = bch2_target_to_mask(c, target) ?: + &c->rw_devs[BCH_DATA_user]; + for_each_set_bit(d, devs->d, BCH_SB_MEMBERS_MAX) { ca = rcu_dereference(c->devs[d]); if (!ca) @@ -106,7 +121,7 @@ void bch2_latency_acct(struct bch_dev *ca, u64 submit_time, int rw) * the time: */ if (abs((int) (old - io_latency)) < (old >> 1) && - now & ~(~0 << 5)) + now & ~(~0U << 5)) break; new = ewma_add(old, io_latency, 5); @@ -121,23 +136,22 @@ void bch2_latency_acct(struct bch_dev *ca, u64 submit_time, int rw) void bch2_bio_free_pages_pool(struct bch_fs *c, struct bio *bio) { + struct bvec_iter_all iter; struct bio_vec *bv; - unsigned i; - bio_for_each_segment_all(bv, bio, i) + bio_for_each_segment_all(bv, bio, iter) if (bv->bv_page != ZERO_PAGE(0)) mempool_free(bv->bv_page, &c->bio_bounce_pages); bio->bi_vcnt = 0; } -static void bch2_bio_alloc_page_pool(struct bch_fs *c, struct bio *bio, - bool *using_mempool) +static struct page *__bio_alloc_page_pool(struct bch_fs *c, bool *using_mempool) { - struct bio_vec *bv = &bio->bi_io_vec[bio->bi_vcnt++]; + struct page *page; if (likely(!*using_mempool)) { - bv->bv_page = alloc_page(GFP_NOIO); - if (unlikely(!bv->bv_page)) { + page = alloc_page(GFP_NOIO); + if (unlikely(!page)) { mutex_lock(&c->bio_bounce_pages_lock); *using_mempool = true; goto pool_alloc; @@ -145,55 +159,295 @@ static void bch2_bio_alloc_page_pool(struct bch_fs *c, struct bio *bio, } } else { pool_alloc: - bv->bv_page = mempool_alloc(&c->bio_bounce_pages, GFP_NOIO); + page = mempool_alloc(&c->bio_bounce_pages, GFP_NOIO); } - bv->bv_len = PAGE_SIZE; - bv->bv_offset = 0; + return page; } void bch2_bio_alloc_pages_pool(struct bch_fs *c, struct bio *bio, - size_t bytes) + size_t size) { bool using_mempool = false; - BUG_ON(DIV_ROUND_UP(bytes, PAGE_SIZE) > bio->bi_max_vecs); + while (size) { + struct page *page = __bio_alloc_page_pool(c, &using_mempool); + unsigned len = min_t(size_t, PAGE_SIZE, size); - bio->bi_iter.bi_size = bytes; - - while (bio->bi_vcnt < DIV_ROUND_UP(bytes, PAGE_SIZE)) - bch2_bio_alloc_page_pool(c, bio, &using_mempool); + BUG_ON(!bio_add_page(bio, page, len, 0)); + size -= len; + } if (using_mempool) mutex_unlock(&c->bio_bounce_pages_lock); } -void bch2_bio_alloc_more_pages_pool(struct bch_fs *c, struct bio *bio, - size_t bytes) +/* Extent update path: */ + +int bch2_sum_sector_overwrites(struct btree_trans *trans, + struct btree_iter *extent_iter, + struct bkey_i *new, + bool *usage_increasing, + s64 *i_sectors_delta, + s64 *disk_sectors_delta) { - while (bio->bi_vcnt < DIV_ROUND_UP(bytes, PAGE_SIZE)) { - struct bio_vec *bv = &bio->bi_io_vec[bio->bi_vcnt]; + struct bch_fs *c = trans->c; + struct btree_iter iter; + struct bkey_s_c old; + unsigned new_replicas = bch2_bkey_replicas(c, bkey_i_to_s_c(new)); + bool new_compressed = bch2_bkey_sectors_compressed(bkey_i_to_s_c(new)); + int ret = 0; - BUG_ON(bio->bi_vcnt >= bio->bi_max_vecs); + *usage_increasing = false; + *i_sectors_delta = 0; + *disk_sectors_delta = 0; - bv->bv_page = alloc_page(GFP_NOIO); - if (!bv->bv_page) { - /* - * We already allocated from mempool, we can't allocate from it again - * without freeing the pages we already allocated or else we could - * deadlock: - */ - bch2_bio_free_pages_pool(c, bio); - bch2_bio_alloc_pages_pool(c, bio, bytes); - return; + bch2_trans_copy_iter(&iter, extent_iter); + + for_each_btree_key_continue_norestart(iter, BTREE_ITER_SLOTS, old, ret) { + s64 sectors = min(new->k.p.offset, old.k->p.offset) - + max(bkey_start_offset(&new->k), + bkey_start_offset(old.k)); + + *i_sectors_delta += sectors * + (bkey_extent_is_allocation(&new->k) - + bkey_extent_is_allocation(old.k)); + + *disk_sectors_delta += sectors * bch2_bkey_nr_ptrs_allocated(bkey_i_to_s_c(new)); + *disk_sectors_delta -= new->k.p.snapshot == old.k->p.snapshot + ? sectors * bch2_bkey_nr_ptrs_fully_allocated(old) + : 0; + + if (!*usage_increasing && + (new->k.p.snapshot != old.k->p.snapshot || + new_replicas > bch2_bkey_replicas(c, old) || + (!new_compressed && bch2_bkey_sectors_compressed(old)))) + *usage_increasing = true; + + if (bkey_cmp(old.k->p, new->k.p) >= 0) + break; + } + + bch2_trans_iter_exit(trans, &iter); + return ret; +} + +int bch2_extent_update(struct btree_trans *trans, + subvol_inum inum, + struct btree_iter *iter, + struct bkey_i *k, + struct disk_reservation *disk_res, + u64 *journal_seq, + u64 new_i_size, + s64 *i_sectors_delta_total, + bool check_enospc) +{ + struct btree_iter inode_iter; + struct bch_inode_unpacked inode_u; + struct bpos next_pos; + bool usage_increasing; + s64 i_sectors_delta = 0, disk_sectors_delta = 0; + int ret; + + /* + * This traverses us the iterator without changing iter->path->pos to + * search_key() (which is pos + 1 for extents): we want there to be a + * path already traversed at iter->pos because + * bch2_trans_extent_update() will use it to attempt extent merging + */ + ret = __bch2_btree_iter_traverse(iter); + if (ret) + return ret; + + ret = bch2_extent_trim_atomic(trans, iter, k); + if (ret) + return ret; + + new_i_size = min(k->k.p.offset << 9, new_i_size); + next_pos = k->k.p; + + ret = bch2_sum_sector_overwrites(trans, iter, k, + &usage_increasing, + &i_sectors_delta, + &disk_sectors_delta); + if (ret) + return ret; + + if (disk_res && + disk_sectors_delta > (s64) disk_res->sectors) { + ret = bch2_disk_reservation_add(trans->c, disk_res, + disk_sectors_delta - disk_res->sectors, + !check_enospc || !usage_increasing + ? BCH_DISK_RESERVATION_NOFAIL : 0); + if (ret) + return ret; + } + + ret = bch2_inode_peek(trans, &inode_iter, &inode_u, inum, + BTREE_ITER_INTENT); + if (ret) + return ret; + + if (!(inode_u.bi_flags & BCH_INODE_I_SIZE_DIRTY) && + new_i_size > inode_u.bi_size) + inode_u.bi_size = new_i_size; + + inode_u.bi_sectors += i_sectors_delta; + + ret = bch2_trans_update(trans, iter, k, 0) ?: + bch2_inode_write(trans, &inode_iter, &inode_u) ?: + bch2_trans_commit(trans, disk_res, journal_seq, + BTREE_INSERT_NOCHECK_RW| + BTREE_INSERT_NOFAIL); + bch2_trans_iter_exit(trans, &inode_iter); + + if (ret) + return ret; + + if (i_sectors_delta_total) + *i_sectors_delta_total += i_sectors_delta; + bch2_btree_iter_set_pos(iter, next_pos); + + return 0; +} + +/* + * Returns -EINTR if we had to drop locks: + */ +int bch2_fpunch_at(struct btree_trans *trans, struct btree_iter *iter, + subvol_inum inum, u64 end, + s64 *i_sectors_delta) +{ + struct bch_fs *c = trans->c; + unsigned max_sectors = KEY_SIZE_MAX & (~0 << c->block_bits); + struct bpos end_pos = POS(inum.inum, end); + struct bkey_s_c k; + int ret = 0, ret2 = 0; + u32 snapshot; + + while (!ret || ret == -EINTR) { + struct disk_reservation disk_res = + bch2_disk_reservation_init(c, 0); + struct bkey_i delete; + + if (ret) + ret2 = ret; + + bch2_trans_begin(trans); + + ret = bch2_subvolume_get_snapshot(trans, inum.subvol, &snapshot); + if (ret) + continue; + + bch2_btree_iter_set_snapshot(iter, snapshot); + + k = bch2_btree_iter_peek(iter); + if (bkey_cmp(iter->pos, end_pos) >= 0) { + bch2_btree_iter_set_pos(iter, end_pos); + break; } - bv->bv_len = PAGE_SIZE; - bv->bv_offset = 0; - bio->bi_vcnt++; + ret = bkey_err(k); + if (ret) + continue; + + bkey_init(&delete.k); + delete.k.p = iter->pos; + + /* create the biggest key we can */ + bch2_key_resize(&delete.k, max_sectors); + bch2_cut_back(end_pos, &delete); + + ret = bch2_extent_update(trans, inum, iter, &delete, + &disk_res, NULL, + 0, i_sectors_delta, false); + bch2_disk_reservation_put(c, &disk_res); } - bio->bi_iter.bi_size = bytes; + return ret ?: ret2; +} + +int bch2_fpunch(struct bch_fs *c, subvol_inum inum, u64 start, u64 end, + s64 *i_sectors_delta) +{ + struct btree_trans trans; + struct btree_iter iter; + int ret; + + bch2_trans_init(&trans, c, BTREE_ITER_MAX, 1024); + bch2_trans_iter_init(&trans, &iter, BTREE_ID_extents, + POS(inum.inum, start), + BTREE_ITER_INTENT); + + ret = bch2_fpunch_at(&trans, &iter, inum, end, i_sectors_delta); + + bch2_trans_iter_exit(&trans, &iter); + bch2_trans_exit(&trans); + + return ret == -EINTR ? 0 : ret; +} + +int bch2_write_index_default(struct bch_write_op *op) +{ + struct bch_fs *c = op->c; + struct bkey_buf sk; + struct open_bucket *ec_ob = ec_open_bucket(c, &op->open_buckets); + struct keylist *keys = &op->insert_keys; + struct bkey_i *k = bch2_keylist_front(keys); + struct btree_trans trans; + struct btree_iter iter; + subvol_inum inum = { + .subvol = op->subvol, + .inum = k->k.p.inode, + }; + int ret; + + BUG_ON(!inum.subvol); + + bch2_bkey_buf_init(&sk); + bch2_trans_init(&trans, c, BTREE_ITER_MAX, 1024); + + do { + bch2_trans_begin(&trans); + + k = bch2_keylist_front(keys); + bch2_bkey_buf_copy(&sk, c, k); + + ret = bch2_subvolume_get_snapshot(&trans, inum.subvol, + &sk.k->k.p.snapshot); + if (ret == -EINTR) + continue; + if (ret) + break; + + bch2_trans_iter_init(&trans, &iter, BTREE_ID_extents, + bkey_start_pos(&sk.k->k), + BTREE_ITER_SLOTS|BTREE_ITER_INTENT); + + ret = bch2_extent_update(&trans, inum, &iter, sk.k, + &op->res, op_journal_seq(op), + op->new_i_size, &op->i_sectors_delta, + op->flags & BCH_WRITE_CHECK_ENOSPC); + bch2_trans_iter_exit(&trans, &iter); + + if (ret == -EINTR) + continue; + if (ret) + break; + + if (ec_ob) + bch2_ob_add_backpointer(c, ec_ob, &sk.k->k); + + if (bkey_cmp(iter.pos, k->k.p) >= 0) + bch2_keylist_pop_front(&op->insert_keys); + else + bch2_cut_front(iter.pos, k); + } while (!bch2_keylist_empty(keys)); + + bch2_trans_exit(&trans); + bch2_bkey_buf_exit(&sk, c); + + return ret; } /* Writes */ @@ -202,20 +456,20 @@ void bch2_submit_wbio_replicas(struct bch_write_bio *wbio, struct bch_fs *c, enum bch_data_type type, const struct bkey_i *k) { - struct bkey_s_c_extent e = bkey_i_to_s_c_extent(k); + struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(bkey_i_to_s_c(k)); const struct bch_extent_ptr *ptr; struct bch_write_bio *n; struct bch_dev *ca; BUG_ON(c->opts.nochanges); - extent_for_each_ptr(e, ptr) { + bkey_for_each_ptr(ptrs, ptr) { BUG_ON(ptr->dev >= BCH_SB_MEMBERS_MAX || !c->devs[ptr->dev]); ca = bch_dev_bkey_exists(c, ptr->dev); - if (ptr + 1 < &extent_entry_last(e)->ptr) { + if (to_entry(ptr + 1) < ptrs.end) { n = to_wbio(bio_clone_fast(&wbio->bio, GFP_NOIO, &ca->replica_set)); @@ -234,13 +488,11 @@ void bch2_submit_wbio_replicas(struct bch_write_bio *wbio, struct bch_fs *c, n->c = c; n->dev = ptr->dev; - n->have_ioref = bch2_dev_get_ioref(ca, WRITE); + n->have_ioref = bch2_dev_get_ioref(ca, + type == BCH_DATA_btree ? READ : WRITE); n->submit_time = local_clock(); n->bio.bi_iter.bi_sector = ptr->offset; - if (!journal_flushes_device(ca)) - n->bio.bi_opf |= REQ_FUA; - if (likely(n->have_ioref)) { this_cpu_add(ca->io_done->sectors[WRITE][type], bio_sectors(&n->bio)); @@ -264,33 +516,19 @@ static void bch2_write_done(struct closure *cl) if (!op->error && (op->flags & BCH_WRITE_FLUSH)) op->error = bch2_journal_error(&c->journal); - if (!(op->flags & BCH_WRITE_NOPUT_RESERVATION)) - bch2_disk_reservation_put(c, &op->res); + bch2_disk_reservation_put(c, &op->res); percpu_ref_put(&c->writes); bch2_keylist_free(&op->insert_keys, op->inline_keys); bch2_time_stats_update(&c->times[BCH_TIME_data_write], op->start_time); - closure_return(cl); -} - -int bch2_write_index_default(struct bch_write_op *op) -{ - struct keylist *keys = &op->insert_keys; - struct btree_iter iter; - int ret; - - bch2_btree_iter_init(&iter, op->c, BTREE_ID_EXTENTS, - bkey_start_pos(&bch2_keylist_front(keys)->k), - BTREE_ITER_INTENT); - - ret = bch2_btree_insert_list_at(&iter, keys, &op->res, - op_journal_seq(op), - BTREE_INSERT_NOFAIL| - BTREE_INSERT_USE_RESERVE); - bch2_btree_iter_unlock(&iter); - - return ret; + if (op->end_io) { + EBUG_ON(cl->parent); + closure_debug_destroy(cl); + op->end_io(op); + } else { + closure_return(cl); + } } /** @@ -300,31 +538,26 @@ static void __bch2_write_index(struct bch_write_op *op) { struct bch_fs *c = op->c; struct keylist *keys = &op->insert_keys; - struct bkey_s_extent e; struct bch_extent_ptr *ptr; struct bkey_i *src, *dst = keys->keys, *n, *k; + unsigned dev; int ret; for (src = keys->keys; src != keys->top; src = n) { n = bkey_next(src); - bkey_copy(dst, src); - e = bkey_i_to_s_extent(dst); - extent_for_each_ptr_backwards(e, ptr) - if (test_bit(ptr->dev, op->failed.d)) - bch2_extent_drop_ptr(e, ptr); + if (bkey_extent_is_direct_data(&src->k)) { + bch2_bkey_drop_ptrs(bkey_i_to_s(src), ptr, + test_bit(ptr->dev, op->failed.d)); - if (!bch2_extent_nr_ptrs(e.c)) { - ret = -EIO; - goto err; - } - - if (!(op->flags & BCH_WRITE_NOMARK_REPLICAS)) { - ret = bch2_mark_bkey_replicas(c, BCH_DATA_USER, e.s_c); - if (ret) + if (!bch2_bkey_nr_ptrs(bkey_i_to_s_c(src))) { + ret = -EIO; goto err; + } } + if (dst != src) + memmove_u64s_down(dst, src, src->u64s); dst = bkey_next(dst); } @@ -335,23 +568,34 @@ static void __bch2_write_index(struct bch_write_op *op) * particularly want to plumb io_opts all the way through the btree * update stack right now */ - for_each_keylist_key(keys, k) + for_each_keylist_key(keys, k) { bch2_rebalance_add_key(c, bkey_i_to_s_c(k), &op->opts); + if (bch2_bkey_is_incompressible(bkey_i_to_s_c(k))) + bch2_check_set_feature(op->c, BCH_FEATURE_incompressible); + + } + if (!bch2_keylist_empty(keys)) { u64 sectors_start = keylist_sectors(keys); int ret = op->index_update_fn(op); + BUG_ON(ret == -EINTR); BUG_ON(keylist_sectors(keys) && !ret); op->written += sectors_start - keylist_sectors(keys); if (ret) { - __bcache_io_error(c, "btree IO error %i", ret); + bch_err_inum_ratelimited(c, op->pos.inode, + "write error %i from btree update", ret); op->error = ret; } } out: + /* If some a bucket wasn't written, we can't erasure code it: */ + for_each_set_bit(dev, op->failed.d, BCH_SB_MEMBERS_MAX) + bch2_open_bucket_write_error(c, &op->open_buckets, dev); + bch2_open_buckets_put(c, &op->open_buckets); return; err: @@ -367,7 +611,9 @@ static void bch2_write_index(struct closure *cl) __bch2_write_index(op); - if (!op->error && (op->flags & BCH_WRITE_FLUSH)) { + if (!(op->flags & BCH_WRITE_DONE)) { + continue_at(cl, __bch2_write, index_update_wq(op)); + } else if (!op->error && (op->flags & BCH_WRITE_FLUSH)) { bch2_journal_flush_seq_async(&c->journal, *op_journal_seq(op), cl); @@ -386,7 +632,11 @@ static void bch2_write_endio(struct bio *bio) struct bch_fs *c = wbio->c; struct bch_dev *ca = bch_dev_bkey_exists(c, wbio->dev); - if (bch2_dev_io_err_on(bio->bi_status, ca, "data write")) + if (bch2_dev_inum_io_err_on(bio->bi_status, ca, + op->pos.inode, + op->pos.offset - bio_sectors(bio), /* XXX definitely wrong */ + "data write error: %s", + bch2_blk_status_to_str(bio->bi_status))) set_bit(wbio->dev, op->failed.d); if (wbio->have_ioref) { @@ -402,8 +652,10 @@ static void bch2_write_endio(struct bio *bio) if (parent) bio_endio(&parent->bio); - else + else if (!(op->flags & BCH_WRITE_SKIP_CLOSURE_PUT)) closure_put(cl); + else + continue_at_nobarrier(cl, bch2_write_index, index_update_wq(op)); } static void init_append_extent(struct bch_write_op *op, @@ -411,16 +663,23 @@ static void init_append_extent(struct bch_write_op *op, struct bversion version, struct bch_extent_crc_unpacked crc) { - struct bkey_i_extent *e = bkey_extent_init(op->insert_keys.top); + struct bch_fs *c = op->c; + struct bkey_i_extent *e; op->pos.offset += crc.uncompressed_size; - e->k.p = op->pos; - e->k.size = crc.uncompressed_size; - e->k.version = version; - bkey_extent_set_cached(&e->k, op->flags & BCH_WRITE_CACHED); - bch2_extent_crc_append(e, crc); - bch2_alloc_sectors_append_ptrs(op->c, wp, e, crc.compressed_size); + e = bkey_extent_init(op->insert_keys.top); + e->k.p = op->pos; + e->k.size = crc.uncompressed_size; + e->k.version = version; + + if (crc.csum_type || + crc.compression_type || + crc.nonce) + bch2_extent_crc_append(&e->k_i, crc); + + bch2_alloc_sectors_append_ptrs(c, wp, &e->k_i, crc.compressed_size, + op->flags & BCH_WRITE_CACHED); bch2_keylist_push(&op->insert_keys); } @@ -428,50 +687,48 @@ static void init_append_extent(struct bch_write_op *op, static struct bio *bch2_write_bio_alloc(struct bch_fs *c, struct write_point *wp, struct bio *src, - bool *page_alloc_failed) + bool *page_alloc_failed, + void *buf) { struct bch_write_bio *wbio; struct bio *bio; unsigned output_available = min(wp->sectors_free << 9, src->bi_iter.bi_size); - unsigned pages = DIV_ROUND_UP(output_available, PAGE_SIZE); + unsigned pages = DIV_ROUND_UP(output_available + + (buf + ? ((unsigned long) buf & (PAGE_SIZE - 1)) + : 0), PAGE_SIZE); + + pages = min(pages, BIO_MAX_VECS); bio = bio_alloc_bioset(GFP_NOIO, pages, &c->bio_write); wbio = wbio_init(bio); - wbio->bounce = true; wbio->put_bio = true; /* copy WRITE_SYNC flag */ wbio->bio.bi_opf = src->bi_opf; + if (buf) { + bch2_bio_map(bio, buf, output_available); + return bio; + } + + wbio->bounce = true; + /* * We can't use mempool for more than c->sb.encoded_extent_max * worth of pages, but we'd like to allocate more if we can: */ - while (bio->bi_iter.bi_size < output_available) { - unsigned len = min_t(unsigned, PAGE_SIZE, - output_available - bio->bi_iter.bi_size); - struct page *p; - - p = alloc_page(GFP_NOIO); - if (!p) { - unsigned pool_max = - min_t(unsigned, output_available, - c->sb.encoded_extent_max << 9); - - if (bio_sectors(bio) < pool_max) - bch2_bio_alloc_pages_pool(c, bio, pool_max); - break; - } + bch2_bio_alloc_pages_pool(c, bio, + min_t(unsigned, output_available, + c->opts.encoded_extent_max)); - bio->bi_io_vec[bio->bi_vcnt++] = (struct bio_vec) { - .bv_page = p, - .bv_len = len, - .bv_offset = 0, - }; - bio->bi_iter.bi_size += len; - } + if (bio->bi_iter.bi_size < output_available) + *page_alloc_failed = + bch2_bio_alloc_pages(bio, + output_available - + bio->bi_iter.bi_size, + GFP_NOFS) != 0; - *page_alloc_failed = bio->bi_vcnt < pages; return bio; } @@ -507,6 +764,7 @@ static int bch2_write_decrypt(struct bch_write_op *op) struct bch_fs *c = op->c; struct nonce nonce = extent_nonce(op->version, op->crc); struct bch_csum csum; + int ret; if (!bch2_csum_type_is_encryption(op->crc.csum_type)) return 0; @@ -521,10 +779,10 @@ static int bch2_write_decrypt(struct bch_write_op *op) if (bch2_crc_cmp(op->crc.csum, csum)) return -EIO; - bch2_encrypt_bio(c, op->crc.csum_type, nonce, &op->wbio.bio); + ret = bch2_encrypt_bio(c, op->crc.csum_type, nonce, &op->wbio.bio); op->crc.csum_type = 0; op->crc.csum = (struct bch_csum) { 0, 0 }; - return 0; + return ret; } static enum prep_encoded_ret { @@ -545,8 +803,9 @@ static enum prep_encoded_ret { /* Can we just write the entire extent as is? */ if (op->crc.uncompressed_size == op->crc.live_size && op->crc.compressed_size <= wp->sectors_free && - op->crc.compression_type == op->compression_type) { - if (!op->crc.compression_type && + (op->crc.compression_type == op->compression_type || + op->incompressible)) { + if (!crc_is_compressed(op->crc) && op->csum_type != op->crc.csum_type && bch2_write_rechecksum(c, op, op->csum_type)) return PREP_ENCODED_CHECKSUM_ERR; @@ -558,7 +817,7 @@ static enum prep_encoded_ret { * If the data is compressed and we couldn't write the entire extent as * is, we have to decompress it: */ - if (op->crc.compression_type) { + if (crc_is_compressed(op->crc)) { struct bch_csum csum; if (bch2_write_decrypt(op)) @@ -601,20 +860,22 @@ static enum prep_encoded_ret { return PREP_ENCODED_OK; } -static int bch2_write_extent(struct bch_write_op *op, struct write_point *wp) +static int bch2_write_extent(struct bch_write_op *op, struct write_point *wp, + struct bio **_dst) { struct bch_fs *c = op->c; struct bio *src = &op->wbio.bio, *dst = src; struct bvec_iter saved_iter; - struct bkey_i *key_to_write; - unsigned key_to_write_offset = op->insert_keys.top_p - - op->insert_keys.keys_p; - unsigned total_output = 0; - bool bounce = false, page_alloc_failed = false; + void *ec_buf; + unsigned total_output = 0, total_input = 0; + bool bounce = false; + bool page_alloc_failed = false; int ret, more = 0; BUG_ON(!bio_sectors(src)); + ec_buf = bch2_writepoint_ec_buf(c, wp); + switch (bch2_write_prep_encoded_data(op, wp)) { case PREP_ENCODED_OK: break; @@ -624,16 +885,27 @@ static int bch2_write_extent(struct bch_write_op *op, struct write_point *wp) case PREP_ENCODED_CHECKSUM_ERR: goto csum_err; case PREP_ENCODED_DO_WRITE: + /* XXX look for bug here */ + if (ec_buf) { + dst = bch2_write_bio_alloc(c, wp, src, + &page_alloc_failed, + ec_buf); + bio_copy_data(dst, src); + bounce = true; + } init_append_extent(op, wp, op->version, op->crc); goto do_write; } - if (op->compression_type || + if (ec_buf || + op->compression_type || (op->csum_type && !(op->flags & BCH_WRITE_PAGES_STABLE)) || (bch2_csum_type_is_encryption(op->csum_type) && !(op->flags & BCH_WRITE_PAGES_OWNED))) { - dst = bch2_write_bio_alloc(c, wp, src, &page_alloc_failed); + dst = bch2_write_bio_alloc(c, wp, src, + &page_alloc_failed, + ec_buf); bounce = true; } @@ -646,8 +918,8 @@ static int bch2_write_extent(struct bch_write_op *op, struct write_point *wp) size_t dst_len, src_len; if (page_alloc_failed && - bio_sectors(dst) < wp->sectors_free && - bio_sectors(dst) < c->sb.encoded_extent_max) + dst->bi_iter.bi_size < (wp->sectors_free << 9) && + dst->bi_iter.bi_size < c->opts.encoded_extent_max) break; BUG_ON(op->compression_type && @@ -655,17 +927,19 @@ static int bch2_write_extent(struct bch_write_op *op, struct write_point *wp) bch2_csum_type_is_encryption(op->crc.csum_type)); BUG_ON(op->compression_type && !bounce); - crc.compression_type = op->compression_type - ? bch2_bio_compress(c, dst, &dst_len, src, &src_len, - op->compression_type) + crc.compression_type = op->incompressible + ? BCH_COMPRESSION_TYPE_incompressible + : op->compression_type + ? bch2_bio_compress(c, dst, &dst_len, src, &src_len, + op->compression_type) : 0; - if (!crc.compression_type) { + if (!crc_is_compressed(crc)) { dst_len = min(dst->bi_iter.bi_size, src->bi_iter.bi_size); dst_len = min_t(unsigned, dst_len, wp->sectors_free << 9); if (op->csum_type) dst_len = min_t(unsigned, dst_len, - c->sb.encoded_extent_max << 9); + c->opts.encoded_extent_max); if (bounce) { swap(dst->bi_iter.bi_size, dst_len); @@ -680,7 +954,7 @@ static int bch2_write_extent(struct bch_write_op *op, struct write_point *wp) if (bch2_csum_type_is_encryption(op->csum_type)) { if (bversion_zero(version)) { - version.lo = atomic64_inc_return(&c->key_version) + 1; + version.lo = atomic64_inc_return(&c->key_version); } else { crc.nonce = op->nonce; op->nonce += src_len >> 9; @@ -688,7 +962,7 @@ static int bch2_write_extent(struct bch_write_op *op, struct write_point *wp) } if ((op->flags & BCH_WRITE_DATA_ENCODED) && - !crc.compression_type && + !crc_is_compressed(crc) && bch2_csum_type_is_encryption(op->crc.csum_type) == bch2_csum_type_is_encryption(op->csum_type)) { /* @@ -723,8 +997,11 @@ static int bch2_write_extent(struct bch_write_op *op, struct write_point *wp) crc.live_size = src_len >> 9; swap(dst->bi_iter.bi_size, dst_len); - bch2_encrypt_bio(c, op->csum_type, - extent_nonce(version, crc), dst); + ret = bch2_encrypt_bio(c, op->csum_type, + extent_nonce(version, crc), dst); + if (ret) + goto err; + crc.csum = bch2_checksum_bio(c, op->csum_type, extent_nonce(version, crc), dst); crc.csum_type = op->csum_type; @@ -736,7 +1013,8 @@ static int bch2_write_extent(struct bch_write_op *op, struct write_point *wp) if (dst != src) bio_advance(dst, dst_len); bio_advance(src, src_len); - total_output += dst_len; + total_output += dst_len; + total_input += src_len; } while (dst->bi_iter.bi_size && src->bi_iter.bi_size && wp->sectors_free && @@ -749,42 +1027,29 @@ static int bch2_write_extent(struct bch_write_op *op, struct write_point *wp) dst->bi_iter = saved_iter; - if (!bounce && more) { - dst = bio_split(src, total_output >> 9, + if (dst == src && more) { + BUG_ON(total_output != total_input); + + dst = bio_split(src, total_input >> 9, GFP_NOIO, &c->bio_write); - wbio_init(dst)->put_bio = true; + wbio_init(dst)->put_bio = true; + /* copy WRITE_SYNC flag */ + dst->bi_opf = src->bi_opf; } dst->bi_iter.bi_size = total_output; - - /* Free unneeded pages after compressing: */ - if (bounce) - while (dst->bi_vcnt > DIV_ROUND_UP(dst->bi_iter.bi_size, PAGE_SIZE)) - mempool_free(dst->bi_io_vec[--dst->bi_vcnt].bv_page, - &c->bio_bounce_pages); do_write: - /* might have done a realloc... */ - - key_to_write = (void *) (op->insert_keys.keys_p + key_to_write_offset); - - dst->bi_end_io = bch2_write_endio; - dst->bi_private = &op->cl; - bio_set_op_attrs(dst, REQ_OP_WRITE, 0); - - closure_get(dst->bi_private); - - bch2_submit_wbio_replicas(to_wbio(dst), c, BCH_DATA_USER, - key_to_write); + *_dst = dst; return more; csum_err: bch_err(c, "error verifying existing checksum while " "rewriting existing data (memory corruption?)"); ret = -EIO; err: - if (bounce) { + if (to_wbio(dst)->bounce) bch2_bio_free_pages_pool(c, dst); + if (to_wbio(dst)->put_bio) bio_put(dst); - } return ret; } @@ -794,9 +1059,20 @@ static void __bch2_write(struct closure *cl) struct bch_write_op *op = container_of(cl, struct bch_write_op, cl); struct bch_fs *c = op->c; struct write_point *wp; + struct bio *bio = NULL; + bool skip_put = true; + unsigned nofs_flags; int ret; + + nofs_flags = memalloc_nofs_save(); again: + memset(&op->failed, 0, sizeof(op->failed)); + do { + struct bkey_i *key_to_write; + unsigned key_to_write_offset = op->insert_keys.top_p - + op->insert_keys.keys_p; + /* +1 for possible cache device: */ if (op->open_buckets.nr + op->nr_replicas + 1 > ARRAY_SIZE(op->open_buckets.v)) @@ -808,15 +1084,28 @@ again: BKEY_EXTENT_U64s_MAX)) goto flush_io; + if ((op->flags & BCH_WRITE_FROM_INTERNAL) && + percpu_ref_is_dying(&c->writes)) { + ret = -EROFS; + goto err; + } + + /* + * The copygc thread is now global, which means it's no longer + * freeing up space on specific disks, which means that + * allocations for specific disks may hang arbitrarily long: + */ wp = bch2_alloc_sectors_start(c, op->target, + op->opts.erasure_code && !(op->flags & BCH_WRITE_CACHED), op->write_point, &op->devs_have, op->nr_replicas, op->nr_replicas_required, op->alloc_reserve, op->flags, - (op->flags & BCH_WRITE_ALLOC_NOWAIT) ? NULL : cl); + (op->flags & (BCH_WRITE_ALLOC_NOWAIT| + BCH_WRITE_ONLY_SPECIFIED_DEVS)) ? NULL : cl); EBUG_ON(!wp); if (unlikely(IS_ERR(wp))) { @@ -828,39 +1117,133 @@ again: goto flush_io; } - ret = bch2_write_extent(op, wp); + /* + * It's possible for the allocator to fail, put us on the + * freelist waitlist, and then succeed in one of various retry + * paths: if that happens, we need to disable the skip_put + * optimization because otherwise there won't necessarily be a + * barrier before we free the bch_write_op: + */ + if (atomic_read(&cl->remaining) & CLOSURE_WAITING) + skip_put = false; bch2_open_bucket_get(c, wp, &op->open_buckets); + ret = bch2_write_extent(op, wp, &bio); bch2_alloc_sectors_done(c, wp); if (ret < 0) goto err; + + if (ret) { + skip_put = false; + } else { + /* + * for the skip_put optimization this has to be set + * before we submit the bio: + */ + op->flags |= BCH_WRITE_DONE; + } + + bio->bi_end_io = bch2_write_endio; + bio->bi_private = &op->cl; + bio->bi_opf |= REQ_OP_WRITE; + + if (!skip_put) + closure_get(bio->bi_private); + else + op->flags |= BCH_WRITE_SKIP_CLOSURE_PUT; + + key_to_write = (void *) (op->insert_keys.keys_p + + key_to_write_offset); + + bch2_submit_wbio_replicas(to_wbio(bio), c, BCH_DATA_user, + key_to_write); } while (ret); - continue_at(cl, bch2_write_index, index_update_wq(op)); + if (!skip_put) + continue_at(cl, bch2_write_index, index_update_wq(op)); +out: + memalloc_nofs_restore(nofs_flags); return; err: op->error = ret; + op->flags |= BCH_WRITE_DONE; - continue_at(cl, !bch2_keylist_empty(&op->insert_keys) - ? bch2_write_index - : bch2_write_done, index_update_wq(op)); - return; + continue_at(cl, bch2_write_index, index_update_wq(op)); + goto out; flush_io: + /* + * If the write can't all be submitted at once, we generally want to + * block synchronously as that signals backpressure to the caller. + * + * However, if we're running out of a workqueue, we can't block here + * because we'll be blocking other work items from completing: + */ + if (current->flags & PF_WQ_WORKER) { + continue_at(cl, bch2_write_index, index_update_wq(op)); + goto out; + } + closure_sync(cl); if (!bch2_keylist_empty(&op->insert_keys)) { __bch2_write_index(op); if (op->error) { + op->flags |= BCH_WRITE_DONE; continue_at_nobarrier(cl, bch2_write_done, NULL); - return; + goto out; } } goto again; } +static void bch2_write_data_inline(struct bch_write_op *op, unsigned data_len) +{ + struct closure *cl = &op->cl; + struct bio *bio = &op->wbio.bio; + struct bvec_iter iter; + struct bkey_i_inline_data *id; + unsigned sectors; + int ret; + + bch2_check_set_feature(op->c, BCH_FEATURE_inline_data); + + ret = bch2_keylist_realloc(&op->insert_keys, op->inline_keys, + ARRAY_SIZE(op->inline_keys), + BKEY_U64s + DIV_ROUND_UP(data_len, 8)); + if (ret) { + op->error = ret; + goto err; + } + + sectors = bio_sectors(bio); + op->pos.offset += sectors; + + id = bkey_inline_data_init(op->insert_keys.top); + id->k.p = op->pos; + id->k.version = op->version; + id->k.size = sectors; + + iter = bio->bi_iter; + iter.bi_size = data_len; + memcpy_from_bio(id->v.data, bio, iter); + + while (data_len & 7) + id->v.data[data_len++] = '\0'; + set_bkey_val_bytes(&id->k, data_len); + bch2_keylist_push(&op->insert_keys); + + op->flags |= BCH_WRITE_WROTE_DATA_INLINE; + op->flags |= BCH_WRITE_DONE; + + continue_at_nobarrier(cl, bch2_write_index, NULL); + return; +err: + bch2_write_done(&op->cl); +} + /** * bch_write - handle a write to a cache device or flash only volume * @@ -880,39 +1263,62 @@ flush_io: void bch2_write(struct closure *cl) { struct bch_write_op *op = container_of(cl, struct bch_write_op, cl); + struct bio *bio = &op->wbio.bio; struct bch_fs *c = op->c; + unsigned data_len; BUG_ON(!op->nr_replicas); BUG_ON(!op->write_point.v); BUG_ON(!bkey_cmp(op->pos, POS_MAX)); - BUG_ON(bio_sectors(&op->wbio.bio) > U16_MAX); op->start_time = local_clock(); - - memset(&op->failed, 0, sizeof(op->failed)); - bch2_keylist_init(&op->insert_keys, op->inline_keys); - wbio_init(&op->wbio.bio)->put_bio = false; + wbio_init(bio)->put_bio = false; + + if (bio->bi_iter.bi_size & (c->opts.block_size - 1)) { + bch_err_inum_ratelimited(c, op->pos.inode, + "misaligned write"); + op->error = -EIO; + goto err; + } if (c->opts.nochanges || !percpu_ref_tryget(&c->writes)) { - __bcache_io_error(c, "read only"); op->error = -EROFS; - if (!(op->flags & BCH_WRITE_NOPUT_RESERVATION)) - bch2_disk_reservation_put(c, &op->res); - closure_return(cl); - return; + goto err; } - bch2_increment_clock(c, bio_sectors(&op->wbio.bio), WRITE); + this_cpu_add(c->counters[BCH_COUNTER_io_write], bio_sectors(bio)); + bch2_increment_clock(c, bio_sectors(bio), WRITE); + + data_len = min_t(u64, bio->bi_iter.bi_size, + op->new_i_size - (op->pos.offset << 9)); + + if (c->opts.inline_data && + data_len <= min(block_bytes(c) / 2, 1024U)) { + bch2_write_data_inline(op, data_len); + return; + } continue_at_nobarrier(cl, __bch2_write, NULL); + return; +err: + bch2_disk_reservation_put(c, &op->res); + + if (op->end_io) { + EBUG_ON(cl->parent); + closure_debug_destroy(cl); + op->end_io(op); + } else { + closure_return(cl); + } } /* Cache promotion on read */ struct promote_op { struct closure cl; + struct rcu_head rcu; u64 start_time; struct rhash_head hash; @@ -933,23 +1339,19 @@ static inline bool should_promote(struct bch_fs *c, struct bkey_s_c k, struct bch_io_opts opts, unsigned flags) { - if (!opts.promote_target) - return false; - if (!(flags & BCH_READ_MAY_PROMOTE)) return false; - if (percpu_ref_is_dying(&c->writes)) - return false; - - if (!bkey_extent_is_data(k.k)) + if (!opts.promote_target) return false; - if (bch2_extent_has_target(c, bkey_s_c_to_extent(k), opts.promote_target)) + if (bch2_bkey_has_target(c, k, opts.promote_target)) return false; - if (bch2_target_congested(c, opts.promote_target)) + if (bch2_target_congested(c, opts.promote_target)) { + /* XXX trace this */ return false; + } if (rhashtable_lookup_fast(&c->promote_table, &pos, bch_promote_params)) @@ -966,7 +1368,7 @@ static void promote_free(struct bch_fs *c, struct promote_op *op) bch_promote_params); BUG_ON(ret); percpu_ref_put(&c->writes); - kfree(op); + kfree_rcu(op, rcu); } static void promote_done(struct closure *cl) @@ -1001,31 +1403,28 @@ static void promote_start(struct promote_op *op, struct bch_read_bio *rbio) bch2_migrate_read_done(&op->write, rbio); closure_init(cl, NULL); - closure_call(&op->write.op.cl, bch2_write, c->wq, cl); + closure_call(&op->write.op.cl, bch2_write, c->btree_update_wq, cl); closure_return_with_destructor(cl, promote_done); } -noinline static struct promote_op *__promote_alloc(struct bch_fs *c, + enum btree_id btree_id, + struct bkey_s_c k, struct bpos pos, - struct extent_pick_ptr *pick, + struct extent_ptr_decoded *pick, struct bch_io_opts opts, - unsigned rbio_sectors, + unsigned sectors, struct bch_read_bio **rbio) { struct promote_op *op = NULL; struct bio *bio; - unsigned rbio_pages = DIV_ROUND_UP(rbio_sectors, PAGE_SECTORS); - /* data might have to be decompressed in the write path: */ - unsigned wbio_pages = DIV_ROUND_UP(pick->crc.uncompressed_size, - PAGE_SECTORS); + unsigned pages = DIV_ROUND_UP(sectors, PAGE_SECTORS); int ret; if (!percpu_ref_tryget(&c->writes)) return NULL; - op = kzalloc(sizeof(*op) + sizeof(struct bio_vec) * wbio_pages, - GFP_NOIO); + op = kzalloc(sizeof(*op) + sizeof(struct bio_vec) * pages, GFP_NOIO); if (!op) goto err; @@ -1033,46 +1432,42 @@ static struct promote_op *__promote_alloc(struct bch_fs *c, op->pos = pos; /* - * promotes require bouncing, but if the extent isn't - * checksummed/compressed it might be too big for the mempool: + * We don't use the mempool here because extents that aren't + * checksummed or compressed can be too big for the mempool: */ - if (rbio_sectors > c->sb.encoded_extent_max) { - *rbio = kzalloc(sizeof(struct bch_read_bio) + - sizeof(struct bio_vec) * rbio_pages, - GFP_NOIO); - if (!*rbio) - goto err; + *rbio = kzalloc(sizeof(struct bch_read_bio) + + sizeof(struct bio_vec) * pages, + GFP_NOIO); + if (!*rbio) + goto err; - rbio_init(&(*rbio)->bio, opts); - bio_init(&(*rbio)->bio, (*rbio)->bio.bi_inline_vecs, - rbio_pages); + rbio_init(&(*rbio)->bio, opts); + bio_init(&(*rbio)->bio, (*rbio)->bio.bi_inline_vecs, pages); - (*rbio)->bio.bi_iter.bi_size = rbio_sectors << 9; - bch2_bio_map(&(*rbio)->bio, NULL); + if (bch2_bio_alloc_pages(&(*rbio)->bio, sectors << 9, + GFP_NOIO)) + goto err; - if (bch2_bio_alloc_pages(&(*rbio)->bio, GFP_NOIO)) - goto err; - - (*rbio)->bounce = true; - (*rbio)->split = true; - (*rbio)->kmalloc = true; - } + (*rbio)->bounce = true; + (*rbio)->split = true; + (*rbio)->kmalloc = true; if (rhashtable_lookup_insert_fast(&c->promote_table, &op->hash, bch_promote_params)) goto err; bio = &op->write.op.wbio.bio; - bio_init(bio, bio->bi_inline_vecs, wbio_pages); + bio_init(bio, bio->bi_inline_vecs, pages); ret = bch2_migrate_write_init(c, &op->write, writepoint_hashed((unsigned long) current), opts, DATA_PROMOTE, (struct data_opts) { - .target = opts.promote_target + .target = opts.promote_target, + .nr_replicas = 1, }, - bkey_s_c_null); + btree_id, k); BUG_ON(ret); return op; @@ -1086,10 +1481,11 @@ err: return NULL; } -static inline struct promote_op *promote_alloc(struct bch_fs *c, +noinline +static struct promote_op *promote_alloc(struct bch_fs *c, struct bvec_iter iter, struct bkey_s_c k, - struct extent_pick_ptr *pick, + struct extent_ptr_decoded *pick, struct bch_io_opts opts, unsigned flags, struct bch_read_bio **rbio, @@ -1097,8 +1493,9 @@ static inline struct promote_op *promote_alloc(struct bch_fs *c, bool *read_full) { bool promote_full = *read_full || READ_ONCE(c->promote_whole_extents); + /* data might have to be decompressed in the write path: */ unsigned sectors = promote_full - ? pick->crc.compressed_size + ? max(pick->crc.compressed_size, pick->crc.live_size) : bvec_iter_sectors(iter); struct bpos pos = promote_full ? bkey_start_pos(k.k) @@ -1108,7 +1505,11 @@ static inline struct promote_op *promote_alloc(struct bch_fs *c, if (!should_promote(c, k, pos, opts, flags)) return NULL; - promote = __promote_alloc(c, pos, pick, opts, sectors, rbio); + promote = __promote_alloc(c, + k.k->type == KEY_TYPE_reflink_v + ? BTREE_ID_reflink + : BTREE_ID_extents, + k, pos, pick, opts, sectors, rbio); if (!promote) return NULL; @@ -1174,112 +1575,74 @@ static inline struct bch_read_bio *bch2_rbio_free(struct bch_read_bio *rbio) return rbio; } +/* + * Only called on a top level bch_read_bio to complete an entire read request, + * not a split: + */ static void bch2_rbio_done(struct bch_read_bio *rbio) { - bch2_time_stats_update(&rbio->c->times[BCH_TIME_data_read], - rbio->start_time); + if (rbio->start_time) + bch2_time_stats_update(&rbio->c->times[BCH_TIME_data_read], + rbio->start_time); bio_endio(&rbio->bio); } static void bch2_read_retry_nodecode(struct bch_fs *c, struct bch_read_bio *rbio, - struct bvec_iter bvec_iter, u64 inode, - struct bch_devs_mask *avoid, unsigned flags) + struct bvec_iter bvec_iter, + struct bch_io_failures *failed, + unsigned flags) { + struct btree_trans trans; struct btree_iter iter; - BKEY_PADDED(k) tmp; + struct bkey_buf sk; struct bkey_s_c k; int ret; flags &= ~BCH_READ_LAST_FRAGMENT; + flags |= BCH_READ_MUST_CLONE; + + bch2_bkey_buf_init(&sk); + bch2_trans_init(&trans, c, 0, 0); - bch2_btree_iter_init(&iter, c, BTREE_ID_EXTENTS, - rbio->pos, BTREE_ITER_SLOTS); + bch2_trans_iter_init(&trans, &iter, rbio->data_btree, + rbio->read_pos, BTREE_ITER_SLOTS); retry: rbio->bio.bi_status = 0; k = bch2_btree_iter_peek_slot(&iter); - if (btree_iter_err(k)) { - bch2_btree_iter_unlock(&iter); + if (bkey_err(k)) goto err; - } - bkey_reassemble(&tmp.k, k); - k = bkey_i_to_s_c(&tmp.k); - bch2_btree_iter_unlock(&iter); + bch2_bkey_buf_reassemble(&sk, c, k); + k = bkey_i_to_s_c(sk.k); + bch2_trans_unlock(&trans); - if (!bkey_extent_is_data(k.k) || - !bch2_extent_matches_ptr(c, bkey_i_to_s_c_extent(&tmp.k), - rbio->pick.ptr, - rbio->pos.offset - - rbio->pick.crc.offset)) { + if (!bch2_bkey_matches_ptr(c, k, + rbio->pick.ptr, + rbio->data_pos.offset - + rbio->pick.crc.offset)) { /* extent we wanted to read no longer exists: */ rbio->hole = true; goto out; } - ret = __bch2_read_extent(c, rbio, bvec_iter, k, avoid, flags); + ret = __bch2_read_extent(&trans, rbio, bvec_iter, + rbio->read_pos, + rbio->data_btree, + k, 0, failed, flags); if (ret == READ_RETRY) goto retry; if (ret) goto err; - goto out; -err: - rbio->bio.bi_status = BLK_STS_IOERR; out: bch2_rbio_done(rbio); -} - -static void bch2_read_retry(struct bch_fs *c, struct bch_read_bio *rbio, - struct bvec_iter bvec_iter, u64 inode, - struct bch_devs_mask *avoid, unsigned flags) -{ - struct btree_iter iter; - struct bkey_s_c k; - int ret; - - flags &= ~BCH_READ_LAST_FRAGMENT; - flags |= BCH_READ_MUST_CLONE; -retry: - for_each_btree_key(&iter, c, BTREE_ID_EXTENTS, - POS(inode, bvec_iter.bi_sector), - BTREE_ITER_SLOTS, k) { - BKEY_PADDED(k) tmp; - unsigned bytes; - - bkey_reassemble(&tmp.k, k); - k = bkey_i_to_s_c(&tmp.k); - bch2_btree_iter_unlock(&iter); - - bytes = min_t(unsigned, bvec_iter.bi_size, - (k.k->p.offset - bvec_iter.bi_sector) << 9); - swap(bvec_iter.bi_size, bytes); - - ret = __bch2_read_extent(c, rbio, bvec_iter, k, avoid, flags); - switch (ret) { - case READ_RETRY: - goto retry; - case READ_ERR: - goto err; - }; - - if (bytes == bvec_iter.bi_size) - goto out; - - swap(bvec_iter.bi_size, bytes); - bio_advance_iter(&rbio->bio, &bvec_iter, bytes); - } - - /* - * If we get here, it better have been because there was an error - * reading a btree node - */ - ret = bch2_btree_iter_unlock(&iter); - BUG_ON(!ret); - __bcache_io_error(c, "btree IO error %i", ret); + bch2_trans_iter_exit(&trans, &iter); + bch2_trans_exit(&trans); + bch2_bkey_buf_exit(&sk, c); + return; err: rbio->bio.bi_status = BLK_STS_IOERR; -out: - bch2_rbio_done(rbio); + goto out; } static void bch2_rbio_retry(struct work_struct *work) @@ -1289,15 +1652,16 @@ static void bch2_rbio_retry(struct work_struct *work) struct bch_fs *c = rbio->c; struct bvec_iter iter = rbio->bvec_iter; unsigned flags = rbio->flags; - u64 inode = rbio->pos.inode; - struct bch_devs_mask avoid; + subvol_inum inum = { + .subvol = rbio->subvol, + .inum = rbio->read_pos.inode, + }; + struct bch_io_failures failed = { .nr = 0 }; trace_read_retry(&rbio->bio); - memset(&avoid, 0, sizeof(avoid)); - if (rbio->retry == READ_RETRY_AVOID) - __set_bit(rbio->pick.ptr.dev, avoid.d); + bch2_mark_io_failure(&failed, &rbio->pick); rbio->bio.bi_status = 0; @@ -1306,10 +1670,14 @@ static void bch2_rbio_retry(struct work_struct *work) flags |= BCH_READ_IN_RETRY; flags &= ~BCH_READ_MAY_PROMOTE; - if (flags & BCH_READ_NODECODE) - bch2_read_retry_nodecode(c, rbio, iter, inode, &avoid, flags); - else - bch2_read_retry(c, rbio, iter, inode, &avoid, flags); + if (flags & BCH_READ_NODECODE) { + bch2_read_retry_nodecode(c, rbio, iter, &failed, flags); + } else { + flags &= ~BCH_READ_LAST_FRAGMENT; + flags |= BCH_READ_MUST_CLONE; + + __bch2_read(c, rbio, iter, inum, &failed, flags); + } } static void bch2_rbio_error(struct bch_read_bio *rbio, int retry, @@ -1331,77 +1699,68 @@ static void bch2_rbio_error(struct bch_read_bio *rbio, int retry, } } -static void bch2_rbio_narrow_crcs(struct bch_read_bio *rbio) +static int __bch2_rbio_narrow_crcs(struct btree_trans *trans, + struct bch_read_bio *rbio) { struct bch_fs *c = rbio->c; + u64 data_offset = rbio->data_pos.offset - rbio->pick.crc.offset; + struct bch_extent_crc_unpacked new_crc; struct btree_iter iter; + struct bkey_i *new; struct bkey_s_c k; - struct bkey_i_extent *e; - BKEY_PADDED(k) new; - struct bch_extent_crc_unpacked new_crc; - unsigned offset; - int ret; - - if (rbio->pick.crc.compression_type) - return; + int ret = 0; - bch2_btree_iter_init(&iter, c, BTREE_ID_EXTENTS, rbio->pos, - BTREE_ITER_INTENT); -retry: - k = bch2_btree_iter_peek(&iter); - if (IS_ERR_OR_NULL(k.k)) - goto out; + if (crc_is_compressed(rbio->pick.crc)) + return 0; - if (!bkey_extent_is_data(k.k)) + bch2_trans_iter_init(trans, &iter, rbio->data_btree, rbio->data_pos, + BTREE_ITER_SLOTS|BTREE_ITER_INTENT); + k = bch2_btree_iter_peek_slot(&iter); + if ((ret = bkey_err(k))) goto out; - bkey_reassemble(&new.k, k); - e = bkey_i_to_extent(&new.k); - - if (!bch2_extent_matches_ptr(c, extent_i_to_s_c(e), - rbio->pick.ptr, - rbio->pos.offset - - rbio->pick.crc.offset) || - bversion_cmp(e->k.version, rbio->version)) + if (bversion_cmp(k.k->version, rbio->version) || + !bch2_bkey_matches_ptr(c, k, rbio->pick.ptr, data_offset)) goto out; /* Extent was merged? */ - if (bkey_start_offset(&e->k) < rbio->pos.offset || - e->k.p.offset > rbio->pos.offset + rbio->pick.crc.uncompressed_size) + if (bkey_start_offset(k.k) < data_offset || + k.k->p.offset > data_offset + rbio->pick.crc.uncompressed_size) goto out; - /* The extent might have been partially overwritten since we read it: */ - offset = rbio->pick.crc.offset + (bkey_start_offset(&e->k) - rbio->pos.offset); - if (bch2_rechecksum_bio(c, &rbio->bio, rbio->version, - rbio->pick.crc, NULL, &new_crc, - offset, e->k.size, - rbio->pick.crc.csum_type)) { + rbio->pick.crc, NULL, &new_crc, + bkey_start_offset(k.k) - data_offset, k.k->size, + rbio->pick.crc.csum_type)) { bch_err(c, "error verifying existing checksum while narrowing checksum (memory corruption?)"); + ret = 0; goto out; } - if (!bch2_extent_narrow_crcs(e, new_crc)) + /* + * going to be temporarily appending another checksum entry: + */ + new = bch2_trans_kmalloc(trans, bkey_bytes(k.k) + + sizeof(struct bch_extent_crc128)); + if ((ret = PTR_ERR_OR_ZERO(new))) goto out; - ret = bch2_btree_insert_at(c, NULL, NULL, - BTREE_INSERT_ATOMIC| - BTREE_INSERT_NOFAIL| - BTREE_INSERT_NOWAIT, - BTREE_INSERT_ENTRY(&iter, &e->k_i)); - if (ret == -EINTR) - goto retry; + bkey_reassemble(new, k); + + if (!bch2_bkey_narrow_crcs(new, new_crc)) + goto out; + + ret = bch2_trans_update(trans, &iter, new, + BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE); out: - bch2_btree_iter_unlock(&iter); + bch2_trans_iter_exit(trans, &iter); + return ret; } -static bool should_narrow_crcs(struct bkey_s_c k, - struct extent_pick_ptr *pick, - unsigned flags) +static noinline void bch2_rbio_narrow_crcs(struct bch_read_bio *rbio) { - return !(flags & BCH_READ_IN_RETRY) && - bkey_extent_is_data(k.k) && - bch2_can_narrow_extent_crcs(bkey_s_c_to_extent(k), pick->crc); + bch2_trans_do(rbio->c, NULL, NULL, BTREE_INSERT_NOFAIL, + __bch2_rbio_narrow_crcs(&trans, rbio)); } /* Inner part that may run in process context */ @@ -1416,7 +1775,11 @@ static void __bch2_read_endio(struct work_struct *work) struct bvec_iter dst_iter = rbio->bvec_iter; struct bch_extent_crc_unpacked crc = rbio->pick.crc; struct nonce nonce = extent_nonce(rbio->version, crc); + unsigned nofs_flags; struct bch_csum csum; + int ret; + + nofs_flags = memalloc_nofs_save(); /* Reset iterator for checksumming and copying bounced data: */ if (rbio->bounce) { @@ -1431,6 +1794,13 @@ static void __bch2_read_endio(struct work_struct *work) if (bch2_crc_cmp(csum, rbio->pick.crc.csum)) goto csum_err; + /* + * XXX + * We need to rework the narrow_crcs path to deliver the read completion + * first, and then punt to a different workqueue, otherwise we're + * holding up reads while doing btree updates which is bad for memory + * reclaim. + */ if (unlikely(rbio->narrow_crcs)) bch2_rbio_narrow_crcs(rbio); @@ -1438,11 +1808,14 @@ static void __bch2_read_endio(struct work_struct *work) goto nodecode; /* Adjust crc to point to subset of data we want: */ - crc.offset += rbio->bvec_iter.bi_sector - rbio->pos.offset; + crc.offset += rbio->offset_into_extent; crc.live_size = bvec_iter_sectors(rbio->bvec_iter); - if (crc.compression_type != BCH_COMPRESSION_NONE) { - bch2_encrypt_bio(c, crc.csum_type, nonce, src); + if (crc_is_compressed(crc)) { + ret = bch2_encrypt_bio(c, crc.csum_type, nonce, src); + if (ret) + goto decrypt_err; + if (bch2_bio_uncompress(c, src, dst, dst_iter, crc)) goto decompression_err; } else { @@ -1453,7 +1826,9 @@ static void __bch2_read_endio(struct work_struct *work) BUG_ON(src->bi_iter.bi_size < dst_iter.bi_size); src->bi_iter.bi_size = dst_iter.bi_size; - bch2_encrypt_bio(c, crc.csum_type, nonce, src); + ret = bch2_encrypt_bio(c, crc.csum_type, nonce, src); + if (ret) + goto decrypt_err; if (rbio->bounce) { struct bvec_iter src_iter = src->bi_iter; @@ -1466,7 +1841,10 @@ static void __bch2_read_endio(struct work_struct *work) * Re encrypt data we decrypted, so it's consistent with * rbio->crc: */ - bch2_encrypt_bio(c, crc.csum_type, nonce, src); + ret = bch2_encrypt_bio(c, crc.csum_type, nonce, src); + if (ret) + goto decrypt_err; + promote_start(rbio->promote, rbio); rbio->promote = NULL; } @@ -1475,6 +1853,8 @@ nodecode: rbio = bch2_rbio_free(rbio); bch2_rbio_done(rbio); } +out: + memalloc_nofs_restore(nofs_flags); return; csum_err: /* @@ -1485,22 +1865,25 @@ csum_err: if (!rbio->bounce && (rbio->flags & BCH_READ_USER_MAPPED)) { rbio->flags |= BCH_READ_MUST_BOUNCE; bch2_rbio_error(rbio, READ_RETRY, BLK_STS_IOERR); - return; + goto out; } - bch2_dev_io_error(ca, - "data checksum error, inode %llu offset %llu: expected %0llx:%0llx got %0llx:%0llx (type %u)", - rbio->pos.inode, (u64) rbio->bvec_iter.bi_sector, + bch2_dev_inum_io_error(ca, rbio->read_pos.inode, (u64) rbio->bvec_iter.bi_sector, + "data checksum error: expected %0llx:%0llx got %0llx:%0llx (type %u)", rbio->pick.crc.csum.hi, rbio->pick.crc.csum.lo, csum.hi, csum.lo, crc.csum_type); bch2_rbio_error(rbio, READ_RETRY_AVOID, BLK_STS_IOERR); - return; + goto out; decompression_err: - __bcache_io_error(c, "decompression error, inode %llu offset %llu", - rbio->pos.inode, - (u64) rbio->bvec_iter.bi_sector); + bch_err_inum_ratelimited(c, rbio->read_pos.inode, + "decompression error"); bch2_rbio_error(rbio, READ_ERR, BLK_STS_IOERR); - return; + goto out; +decrypt_err: + bch_err_inum_ratelimited(c, rbio->read_pos.inode, + "decrypt error"); + bch2_rbio_error(rbio, READ_ERR, BLK_STS_IOERR); + goto out; } static void bch2_read_endio(struct bio *bio) @@ -1520,14 +1903,17 @@ static void bch2_read_endio(struct bio *bio) if (!rbio->split) rbio->bio.bi_end_io = rbio->end_io; - if (bch2_dev_io_err_on(bio->bi_status, ca, "data read")) { + if (bch2_dev_inum_io_err_on(bio->bi_status, ca, + rbio->read_pos.inode, + rbio->read_pos.offset, + "data read error: %s", + bch2_blk_status_to_str(bio->bi_status))) { bch2_rbio_error(rbio, READ_RETRY_AVOID, bio->bi_status); return; } - if (rbio->pick.ptr.cached && - (((rbio->flags & BCH_READ_RETRY_IF_STALE) && race_fault()) || - ptr_stale(ca, &rbio->pick.ptr))) { + if (((rbio->flags & BCH_READ_RETRY_IF_STALE) && race_fault()) || + ptr_stale(ca, &rbio->pick.ptr)) { atomic_long_inc(&c->read_realloc_races); if (rbio->flags & BCH_READ_RETRY_IF_STALE) @@ -1538,7 +1924,7 @@ static void bch2_read_endio(struct bio *bio) } if (rbio->narrow_crcs || - rbio->pick.crc.compression_type || + crc_is_compressed(rbio->pick.crc) || bch2_csum_type_is_encryption(rbio->pick.crc.csum_type)) context = RBIO_CONTEXT_UNBOUND, wq = system_unbound_wq; else if (rbio->pick.crc.csum_type) @@ -1547,29 +1933,140 @@ static void bch2_read_endio(struct bio *bio) bch2_rbio_punt(rbio, __bch2_read_endio, context, wq); } -int __bch2_read_extent(struct bch_fs *c, struct bch_read_bio *orig, - struct bvec_iter iter, struct bkey_s_c k, - struct bch_devs_mask *avoid, unsigned flags) +int __bch2_read_indirect_extent(struct btree_trans *trans, + unsigned *offset_into_extent, + struct bkey_buf *orig_k) +{ + struct btree_iter iter; + struct bkey_s_c k; + u64 reflink_offset; + int ret; + + reflink_offset = le64_to_cpu(bkey_i_to_reflink_p(orig_k->k)->v.idx) + + *offset_into_extent; + + bch2_trans_iter_init(trans, &iter, BTREE_ID_reflink, + POS(0, reflink_offset), + BTREE_ITER_SLOTS); + k = bch2_btree_iter_peek_slot(&iter); + ret = bkey_err(k); + if (ret) + goto err; + + if (k.k->type != KEY_TYPE_reflink_v && + k.k->type != KEY_TYPE_indirect_inline_data) { + bch_err_inum_ratelimited(trans->c, orig_k->k->k.p.inode, + "%llu len %u points to nonexistent indirect extent %llu", + orig_k->k->k.p.offset, + orig_k->k->k.size, + reflink_offset); + bch2_inconsistent_error(trans->c); + ret = -EIO; + goto err; + } + + *offset_into_extent = iter.pos.offset - bkey_start_offset(k.k); + bch2_bkey_buf_reassemble(orig_k, trans->c, k); +err: + bch2_trans_iter_exit(trans, &iter); + return ret; +} + +static noinline void read_from_stale_dirty_pointer(struct btree_trans *trans, + struct bkey_s_c k, + struct bch_extent_ptr ptr) +{ + struct bch_fs *c = trans->c; + struct bch_dev *ca = bch_dev_bkey_exists(c, ptr.dev); + struct btree_iter iter; + struct printbuf buf = PRINTBUF; + int ret; + + bch2_trans_iter_init(trans, &iter, BTREE_ID_alloc, + PTR_BUCKET_POS(c, &ptr), + BTREE_ITER_CACHED); + + pr_buf(&buf, "Attempting to read from stale dirty pointer:"); + pr_indent_push(&buf, 2); + pr_newline(&buf); + + bch2_bkey_val_to_text(&buf, c, k); + pr_newline(&buf); + + pr_buf(&buf, "memory gen: %u", *bucket_gen(ca, iter.pos.offset)); + + ret = lockrestart_do(trans, bkey_err(k = bch2_btree_iter_peek_slot(&iter))); + if (!ret) { + pr_newline(&buf); + bch2_bkey_val_to_text(&buf, c, k); + } + + bch2_fs_inconsistent(c, "%s", buf.buf); + + bch2_trans_iter_exit(trans, &iter); + printbuf_exit(&buf); +} + +int __bch2_read_extent(struct btree_trans *trans, struct bch_read_bio *orig, + struct bvec_iter iter, struct bpos read_pos, + enum btree_id data_btree, struct bkey_s_c k, + unsigned offset_into_extent, + struct bch_io_failures *failed, unsigned flags) { - struct extent_pick_ptr pick; + struct bch_fs *c = trans->c; + struct extent_ptr_decoded pick; struct bch_read_bio *rbio = NULL; - struct bch_dev *ca; + struct bch_dev *ca = NULL; struct promote_op *promote = NULL; bool bounce = false, read_full = false, narrow_crcs = false; - struct bpos pos = bkey_start_pos(k.k); + struct bpos data_pos = bkey_start_pos(k.k); int pick_ret; - pick_ret = bch2_extent_pick_ptr(c, k, avoid, &pick); + if (bkey_extent_is_inline_data(k.k)) { + unsigned bytes = min_t(unsigned, iter.bi_size, + bkey_inline_data_bytes(k.k)); + + swap(iter.bi_size, bytes); + memcpy_to_bio(&orig->bio, iter, bkey_inline_data_p(k)); + swap(iter.bi_size, bytes); + bio_advance_iter(&orig->bio, &iter, bytes); + zero_fill_bio_iter(&orig->bio, iter); + goto out_read_done; + } +retry_pick: + pick_ret = bch2_bkey_pick_read_device(c, k, failed, &pick); /* hole or reservation - just zero fill: */ if (!pick_ret) goto hole; - if (pick_ret < 0) - goto no_device; + if (pick_ret < 0) { + bch_err_inum_ratelimited(c, k.k->p.inode, + "no device to read from"); + goto err; + } - if (pick_ret > 0) - ca = bch_dev_bkey_exists(c, pick.ptr.dev); + ca = bch_dev_bkey_exists(c, pick.ptr.dev); + + /* + * Stale dirty pointers are treated as IO errors, but @failed isn't + * allocated unless we're in the retry path - so if we're not in the + * retry path, don't check here, it'll be caught in bch2_read_endio() + * and we'll end up in the retry path: + */ + if ((flags & BCH_READ_IN_RETRY) && + !pick.ptr.cached && + unlikely(ptr_stale(ca, &pick.ptr))) { + read_from_stale_dirty_pointer(trans, k, pick.ptr); + bch2_mark_io_failure(failed, &pick); + goto retry_pick; + } + + /* + * Unlock the iterator while the btree node's lock is still in + * cache, before doing the IO: + */ + bch2_trans_unlock(trans); if (flags & BCH_READ_NODECODE) { /* @@ -1579,25 +2076,24 @@ int __bch2_read_extent(struct bch_fs *c, struct bch_read_bio *orig, if (pick.crc.compressed_size > orig->bio.bi_vcnt * PAGE_SECTORS) goto hole; - iter.bi_sector = pos.offset; iter.bi_size = pick.crc.compressed_size << 9; - goto noclone; + goto get_bio; } if (!(flags & BCH_READ_LAST_FRAGMENT) || bio_flagged(&orig->bio, BIO_CHAIN)) flags |= BCH_READ_MUST_CLONE; - narrow_crcs = should_narrow_crcs(k, &pick, flags); + narrow_crcs = !(flags & BCH_READ_IN_RETRY) && + bch2_can_narrow_extent_crcs(k, pick.crc); if (narrow_crcs && (flags & BCH_READ_USER_MAPPED)) flags |= BCH_READ_MUST_BOUNCE; - EBUG_ON(bkey_start_offset(k.k) > iter.bi_sector || - k.k->p.offset < bvec_iter_end_sector(iter)); + EBUG_ON(offset_into_extent + bvec_iter_sectors(iter) > k.k->size); - if (pick.crc.compression_type != BCH_COMPRESSION_NONE || - (pick.crc.csum_type != BCH_CSUM_NONE && + if (crc_is_compressed(pick.crc) || + (pick.crc.csum_type != BCH_CSUM_none && (bvec_iter_sectors(iter) != pick.crc.uncompressed_size || (bch2_csum_type_is_encryption(pick.crc.csum_type) && (flags & BCH_READ_USER_MAPPED)) || @@ -1606,28 +2102,40 @@ int __bch2_read_extent(struct bch_fs *c, struct bch_read_bio *orig, bounce = true; } - promote = promote_alloc(c, iter, k, &pick, orig->opts, flags, - &rbio, &bounce, &read_full); + if (orig->opts.promote_target) + promote = promote_alloc(c, iter, k, &pick, orig->opts, flags, + &rbio, &bounce, &read_full); if (!read_full) { - EBUG_ON(pick.crc.compression_type); + EBUG_ON(crc_is_compressed(pick.crc)); EBUG_ON(pick.crc.csum_type && (bvec_iter_sectors(iter) != pick.crc.uncompressed_size || bvec_iter_sectors(iter) != pick.crc.live_size || pick.crc.offset || - iter.bi_sector != pos.offset)); + offset_into_extent)); + data_pos.offset += offset_into_extent; pick.ptr.offset += pick.crc.offset + - (iter.bi_sector - pos.offset); + offset_into_extent; + offset_into_extent = 0; pick.crc.compressed_size = bvec_iter_sectors(iter); pick.crc.uncompressed_size = bvec_iter_sectors(iter); pick.crc.offset = 0; pick.crc.live_size = bvec_iter_sectors(iter); - pos.offset = iter.bi_sector; + offset_into_extent = 0; } - +get_bio: if (rbio) { - /* promote already allocated bounce rbio */ + /* + * promote already allocated bounce rbio: + * promote needs to allocate a bio big enough for uncompressing + * data in the write path, but we're not going to use it all + * here: + */ + EBUG_ON(rbio->bio.bi_iter.bi_size < + pick.crc.compressed_size << 9); + rbio->bio.bi_iter.bi_size = + pick.crc.compressed_size << 9; } else if (bounce) { unsigned sectors = pick.crc.compressed_size; @@ -1654,13 +2162,12 @@ int __bch2_read_extent(struct bch_fs *c, struct bch_read_bio *orig, rbio->bio.bi_iter = iter; rbio->split = true; } else { -noclone: rbio = orig; rbio->bio.bi_iter = iter; - BUG_ON(bio_flagged(&rbio->bio, BIO_CHAIN)); + EBUG_ON(bio_flagged(&rbio->bio, BIO_CHAIN)); } - BUG_ON(bio_sectors(&rbio->bio) != pick.crc.compressed_size); + EBUG_ON(bio_sectors(&rbio->bio) != pick.crc.compressed_size); rbio->c = c; rbio->submit_time = local_clock(); @@ -1669,15 +2176,20 @@ noclone: else rbio->end_io = orig->bio.bi_end_io; rbio->bvec_iter = iter; + rbio->offset_into_extent= offset_into_extent; rbio->flags = flags; rbio->have_ioref = pick_ret > 0 && bch2_dev_get_ioref(ca, READ); rbio->narrow_crcs = narrow_crcs; rbio->hole = 0; rbio->retry = 0; rbio->context = 0; + /* XXX: only initialize this if needed */ rbio->devs_have = bch2_bkey_devs(k); rbio->pick = pick; - rbio->pos = pos; + rbio->subvol = orig->subvol; + rbio->read_pos = read_pos; + rbio->data_btree = data_btree; + rbio->data_pos = data_pos; rbio->version = k.k->version; rbio->promote = promote; INIT_WORK(&rbio->work, NULL); @@ -1689,33 +2201,54 @@ noclone: if (rbio->bounce) trace_read_bounce(&rbio->bio); + this_cpu_add(c->counters[BCH_COUNTER_io_read], bio_sectors(&rbio->bio)); bch2_increment_clock(c, bio_sectors(&rbio->bio), READ); - if (!rbio->have_ioref) - goto no_device_postclone; + /* + * If it's being moved internally, we don't want to flag it as a cache + * hit: + */ + if (pick.ptr.cached && !(flags & BCH_READ_NODECODE)) + bch2_bucket_io_time_reset(trans, pick.ptr.dev, + PTR_BUCKET_NR(ca, &pick.ptr), READ); - percpu_down_read_preempt_disable(&c->usage_lock); - bucket_io_clock_reset(c, ca, PTR_BUCKET_NR(ca, &pick.ptr), READ); - percpu_up_read_preempt_enable(&c->usage_lock); + if (!(flags & (BCH_READ_IN_RETRY|BCH_READ_LAST_FRAGMENT))) { + bio_inc_remaining(&orig->bio); + trace_read_split(&orig->bio); + } - this_cpu_add(ca->io_done->sectors[READ][BCH_DATA_USER], - bio_sectors(&rbio->bio)); + if (!rbio->pick.idx) { + if (!rbio->have_ioref) { + bch_err_inum_ratelimited(c, k.k->p.inode, + "no device to read from"); + bch2_rbio_error(rbio, READ_RETRY_AVOID, BLK_STS_IOERR); + goto out; + } - bio_set_dev(&rbio->bio, ca->disk_sb.bdev); + this_cpu_add(ca->io_done->sectors[READ][BCH_DATA_user], + bio_sectors(&rbio->bio)); + bio_set_dev(&rbio->bio, ca->disk_sb.bdev); - if (likely(!(flags & BCH_READ_IN_RETRY))) { - if (!(flags & BCH_READ_LAST_FRAGMENT)) { - bio_inc_remaining(&orig->bio); - trace_read_split(&orig->bio); + if (likely(!(flags & BCH_READ_IN_RETRY))) + submit_bio(&rbio->bio); + else + submit_bio_wait(&rbio->bio); + } else { + /* Attempting reconstruct read: */ + if (bch2_ec_read_extent(c, rbio)) { + bch2_rbio_error(rbio, READ_RETRY_AVOID, BLK_STS_IOERR); + goto out; } - submit_bio(&rbio->bio); + if (likely(!(flags & BCH_READ_IN_RETRY))) + bio_endio(&rbio->bio); + } +out: + if (likely(!(flags & BCH_READ_IN_RETRY))) { return 0; } else { int ret; - submit_bio_wait(&rbio->bio); - rbio->context = RBIO_CONTEXT_UNBOUND; bch2_read_endio(&rbio->bio); @@ -1723,29 +2256,22 @@ noclone: rbio = bch2_rbio_free(rbio); if (ret == READ_RETRY_AVOID) { - __set_bit(pick.ptr.dev, avoid->d); + bch2_mark_io_failure(failed, &pick); ret = READ_RETRY; } + if (!ret) + goto out_read_done; + return ret; } -no_device_postclone: - if (!rbio->split) - rbio->bio.bi_end_io = rbio->end_io; - bch2_rbio_free(rbio); -no_device: - __bcache_io_error(c, "no device to read from"); - - if (likely(!(flags & BCH_READ_IN_RETRY))) { - orig->bio.bi_status = BLK_STS_IOERR; - - if (flags & BCH_READ_LAST_FRAGMENT) - bch2_rbio_done(orig); - return 0; - } else { +err: + if (flags & BCH_READ_IN_RETRY) return READ_ERR; - } + + orig->bio.bi_status = BLK_STS_IOERR; + goto out_read_done; hole: /* @@ -1757,66 +2283,115 @@ hole: orig->hole = true; zero_fill_bio_iter(&orig->bio, iter); - +out_read_done: if (flags & BCH_READ_LAST_FRAGMENT) bch2_rbio_done(orig); return 0; } -void bch2_read(struct bch_fs *c, struct bch_read_bio *rbio, u64 inode) +void __bch2_read(struct bch_fs *c, struct bch_read_bio *rbio, + struct bvec_iter bvec_iter, subvol_inum inum, + struct bch_io_failures *failed, unsigned flags) { + struct btree_trans trans; struct btree_iter iter; + struct bkey_buf sk; struct bkey_s_c k; - unsigned flags = BCH_READ_RETRY_IF_STALE| - BCH_READ_MAY_PROMOTE| - BCH_READ_USER_MAPPED; + u32 snapshot; int ret; - BUG_ON(rbio->_state); BUG_ON(flags & BCH_READ_NODECODE); - BUG_ON(flags & BCH_READ_IN_RETRY); - rbio->c = c; - rbio->start_time = local_clock(); + bch2_bkey_buf_init(&sk); + bch2_trans_init(&trans, c, 0, 0); +retry: + bch2_trans_begin(&trans); + iter = (struct btree_iter) { NULL }; + + ret = bch2_subvolume_get_snapshot(&trans, inum.subvol, &snapshot); + if (ret) + goto err; + + bch2_trans_iter_init(&trans, &iter, BTREE_ID_extents, + SPOS(inum.inum, bvec_iter.bi_sector, snapshot), + BTREE_ITER_SLOTS); + while (1) { + unsigned bytes, sectors, offset_into_extent; + enum btree_id data_btree = BTREE_ID_extents; + + /* + * read_extent -> io_time_reset may cause a transaction restart + * without returning an error, we need to check for that here: + */ + if (!bch2_trans_relock(&trans)) { + ret = -EINTR; + break; + } + + bch2_btree_iter_set_pos(&iter, + POS(inum.inum, bvec_iter.bi_sector)); + + k = bch2_btree_iter_peek_slot(&iter); + ret = bkey_err(k); + if (ret) + break; - for_each_btree_key(&iter, c, BTREE_ID_EXTENTS, - POS(inode, rbio->bio.bi_iter.bi_sector), - BTREE_ITER_SLOTS, k) { - BKEY_PADDED(k) tmp; - unsigned bytes; + offset_into_extent = iter.pos.offset - + bkey_start_offset(k.k); + sectors = k.k->size - offset_into_extent; + + bch2_bkey_buf_reassemble(&sk, c, k); + + ret = bch2_read_indirect_extent(&trans, &data_btree, + &offset_into_extent, &sk); + if (ret) + break; + + k = bkey_i_to_s_c(sk.k); /* - * Unlock the iterator while the btree node's lock is still in - * cache, before doing the IO: + * With indirect extents, the amount of data to read is the min + * of the original extent and the indirect extent: */ - bkey_reassemble(&tmp.k, k); - k = bkey_i_to_s_c(&tmp.k); - bch2_btree_iter_unlock(&iter); + sectors = min(sectors, k.k->size - offset_into_extent); - bytes = min_t(unsigned, rbio->bio.bi_iter.bi_size, - (k.k->p.offset - rbio->bio.bi_iter.bi_sector) << 9); - swap(rbio->bio.bi_iter.bi_size, bytes); + bytes = min(sectors, bvec_iter_sectors(bvec_iter)) << 9; + swap(bvec_iter.bi_size, bytes); - if (rbio->bio.bi_iter.bi_size == bytes) + if (bvec_iter.bi_size == bytes) flags |= BCH_READ_LAST_FRAGMENT; - bch2_read_extent(c, rbio, k, flags); + ret = __bch2_read_extent(&trans, rbio, bvec_iter, iter.pos, + data_btree, k, + offset_into_extent, failed, flags); + if (ret) + break; if (flags & BCH_READ_LAST_FRAGMENT) - return; + break; - swap(rbio->bio.bi_iter.bi_size, bytes); - bio_advance(&rbio->bio, bytes); + swap(bvec_iter.bi_size, bytes); + bio_advance_iter(&rbio->bio, &bvec_iter, bytes); + + ret = btree_trans_too_many_iters(&trans); + if (ret) + break; } +err: + bch2_trans_iter_exit(&trans, &iter); - /* - * If we get here, it better have been because there was an error - * reading a btree node - */ - ret = bch2_btree_iter_unlock(&iter); - BUG_ON(!ret); - bcache_io_error(c, &rbio->bio, "btree IO error %i", ret); - bch2_rbio_done(rbio); + if (ret == -EINTR || ret == READ_RETRY || ret == READ_RETRY_AVOID) + goto retry; + + bch2_trans_exit(&trans); + bch2_bkey_buf_exit(&sk, c); + + if (ret) { + bch_err_inum_ratelimited(c, inum.inum, + "read error %i from btree lookup", ret); + rbio->bio.bi_status = BLK_STS_IOERR; + bch2_rbio_done(rbio); + } } void bch2_fs_io_exit(struct bch_fs *c) @@ -1840,8 +2415,8 @@ int bch2_fs_io_init(struct bch_fs *c) mempool_init_page_pool(&c->bio_bounce_pages, max_t(unsigned, c->opts.btree_node_size, - c->sb.encoded_extent_max) / - PAGE_SECTORS, 0) || + c->opts.encoded_extent_max) / + PAGE_SIZE, 0) || rhashtable_init(&c->promote_table, &bch_promote_params)) return -ENOMEM;