From b5e8a6992fb1195cb58cb79461ef50f474c27608 Mon Sep 17 00:00:00 2001 From: Kent Overstreet Date: Mon, 2 Nov 2020 23:51:33 -0500 Subject: [PATCH] bcachefs: Improved inode create optimization This shards new inodes into different btree nodes by using the processor ID for the high bits of the new inode number. Much faster than the previous inode create optimization - this also helps with sharding in the other btrees that index by inode number. Signed-off-by: Kent Overstreet Signed-off-by: Kent Overstreet --- fs/bcachefs/bcachefs.h | 7 +-- fs/bcachefs/inode.c | 139 ++++++++++++++--------------------------- fs/bcachefs/super.c | 7 ++- 3 files changed, 54 insertions(+), 99 deletions(-) diff --git a/fs/bcachefs/bcachefs.h b/fs/bcachefs/bcachefs.h index d56057f27e8f0..c14117227dd78 100644 --- a/fs/bcachefs/bcachefs.h +++ b/fs/bcachefs/bcachefs.h @@ -814,11 +814,8 @@ mempool_t bio_bounce_pages; struct mutex verify_lock; #endif - struct mutex inode_create_lock; - unsigned unused_inodes_nr; - u64 unused_inodes[64]; - u32 unused_inodes_gens[64]; - u64 unused_inode_hint; + u64 *unused_inode_hints; + unsigned inode_shard_bits; /* * A btree node on disk could have too many bsets for an iterator to fit diff --git a/fs/bcachefs/inode.c b/fs/bcachefs/inode.c index c55c164be882e..b49c382f5452c 100644 --- a/fs/bcachefs/inode.c +++ b/fs/bcachefs/inode.c @@ -361,55 +361,6 @@ static inline u32 bkey_generation(struct bkey_s_c k) } } -static int scan_free_inums(struct btree_trans *trans) -{ - struct bch_fs *c = trans->c; - struct btree_iter *iter = NULL; - struct bkey_s_c k; - u64 min = BLOCKDEV_INODE_MAX; - u64 max = c->opts.inodes_32bit - ? S32_MAX : S64_MAX; - u64 start = max(min, READ_ONCE(c->unused_inode_hint)); - int ret = 0; - - iter = bch2_trans_get_iter(trans, BTREE_ID_INODES, POS(0, start), - BTREE_ITER_SLOTS); - if (IS_ERR(iter)) - return PTR_ERR(iter); -again: - for_each_btree_key_continue(iter, BTREE_ITER_SLOTS, k, ret) { - if (bkey_cmp(iter->pos, POS(0, max)) > 0) - break; - - /* - * This doesn't check the btree key cache, but we don't care: - * we have to recheck with an intent lock held on the slot we're - * inserting to anyways: - */ - if (k.k->type != KEY_TYPE_inode) { - if (c->unused_inodes_nr < ARRAY_SIZE(c->unused_inodes)) { - c->unused_inodes[c->unused_inodes_nr] = k.k->p.offset; - c->unused_inodes_gens[c->unused_inodes_nr] = bkey_generation(k); - c->unused_inodes_nr++; - } - - if (c->unused_inodes_nr == ARRAY_SIZE(c->unused_inodes)) - goto out; - } - } - - if (!ret && start != min) { - max = start; - start = min; - bch2_btree_iter_set_pos(iter, POS(0, start)); - goto again; - } -out: - c->unused_inode_hint = iter->pos.offset; - bch2_trans_iter_put(trans, iter); - return ret; -} - int bch2_inode_create(struct btree_trans *trans, struct bch_inode_unpacked *inode_u) { @@ -417,64 +368,68 @@ int bch2_inode_create(struct btree_trans *trans, struct bkey_inode_buf *inode_p; struct btree_iter *iter = NULL; struct bkey_s_c k; - u64 inum; - u32 generation; - int ret = 0; + u64 min, max, start, *hint; + int ret; + + unsigned cpu = raw_smp_processor_id(); + unsigned bits = (c->opts.inodes_32bit + ? 31 : 63) - c->inode_shard_bits; + + min = (cpu << bits); + max = (cpu << bits) | ~(ULLONG_MAX << bits); + + min = max_t(u64, min, BLOCKDEV_INODE_MAX); + hint = c->unused_inode_hints + cpu; + + start = READ_ONCE(*hint); + + if (start >= max || start < min) + start = min; inode_p = bch2_trans_kmalloc(trans, sizeof(*inode_p)); if (IS_ERR(inode_p)) return PTR_ERR(inode_p); +again: + for_each_btree_key(trans, iter, BTREE_ID_INODES, POS(0, start), + BTREE_ITER_SLOTS|BTREE_ITER_INTENT, k, ret) { + if (bkey_cmp(iter->pos, POS(0, max)) > 0) + break; - iter = bch2_trans_get_iter(trans, BTREE_ID_INODES, POS_MIN, - BTREE_ITER_CACHED| - BTREE_ITER_INTENT); - if (IS_ERR(iter)) - return PTR_ERR(iter); -retry: - if (!mutex_trylock(&c->inode_create_lock)) { - bch2_trans_unlock(trans); - mutex_lock(&c->inode_create_lock); - if (!bch2_trans_relock(trans)) { - mutex_unlock(&c->inode_create_lock); - ret = -EINTR; - goto err; - } - } - - if (!c->unused_inodes_nr) - ret = scan_free_inums(trans); - if (!ret && !c->unused_inodes_nr) - ret = -ENOSPC; - if (!ret) { - --c->unused_inodes_nr; - inum = c->unused_inodes[c->unused_inodes_nr]; - generation = c->unused_inodes_gens[c->unused_inodes_nr]; + /* + * There's a potential cache coherency issue with the btree key + * cache code here - we're iterating over the btree, skipping + * that cache. We should never see an empty slot that isn't + * actually empty due to a pending update in the key cache + * because the update that creates the inode isn't done with a + * cached iterator, but - better safe than sorry, check the + * cache before using a slot: + */ + if (k.k->type != KEY_TYPE_inode && + !bch2_btree_key_cache_find(c, BTREE_ID_INODES, iter->pos)) + goto found_slot; } - mutex_unlock(&c->inode_create_lock); - - if (ret) - goto err; - - bch2_btree_iter_set_pos(iter, POS(0, inum)); + bch2_trans_iter_put(trans, iter); - /* Recheck that the slot is free with an intent lock held: */ - k = bch2_btree_iter_peek_cached(iter); - ret = bkey_err(k); if (ret) - goto err; + return ret; - if (k.k->type == KEY_TYPE_inode) - goto retry; + if (start != min) { + /* Retry from start */ + start = min; + goto again; + } - inode_u->bi_inum = inum; - inode_u->bi_generation = generation; + return -ENOSPC; +found_slot: + *hint = k.k->p.offset; + inode_u->bi_inum = k.k->p.offset; + inode_u->bi_generation = bkey_generation(k); bch2_inode_pack(inode_p, inode_u); bch2_trans_update(trans, iter, &inode_p->inode.k_i, 0); -err: bch2_trans_iter_put(trans, iter); - return ret; + return 0; } int bch2_inode_rm(struct bch_fs *c, u64 inode_nr) diff --git a/fs/bcachefs/super.c b/fs/bcachefs/super.c index fd78ab2058658..a2ade0df62b5f 100644 --- a/fs/bcachefs/super.c +++ b/fs/bcachefs/super.c @@ -493,6 +493,7 @@ static void __bch2_fs_free(struct bch_fs *c) kfree(c->replicas_gc.entries); kfree(rcu_dereference_protected(c->disk_groups, 1)); kfree(c->journal_seq_blacklist_table); + kfree(c->unused_inode_hints); free_heap(&c->copygc_heap); if (c->journal_reclaim_wq) @@ -703,8 +704,6 @@ static struct bch_fs *bch2_fs_alloc(struct bch_sb *sb, struct bch_opts opts) seqcount_init(&c->usage_lock); - mutex_init(&c->inode_create_lock); - c->copy_gc_enabled = 1; c->rebalance.enabled = 1; c->promote_whole_extents = true; @@ -746,6 +745,8 @@ static struct bch_fs *bch2_fs_alloc(struct bch_sb *sb, struct bch_opts opts) (btree_blocks(c) + 1) * 2 * sizeof(struct sort_iter_set); + c->inode_shard_bits = ilog2(roundup_pow_of_two(num_possible_cpus())); + if (!(c->wq = alloc_workqueue("bcachefs", WQ_FREEZABLE|WQ_MEM_RECLAIM, 1)) || !(c->copygc_wq = alloc_workqueue("bcachefs_copygc", @@ -764,6 +765,8 @@ static struct bch_fs *bch2_fs_alloc(struct bch_sb *sb, struct bch_opts opts) mempool_init_kvpmalloc_pool(&c->btree_bounce_pool, 1, btree_bytes(c)) || mempool_init_kmalloc_pool(&c->large_bkey_pool, 1, 2048) || + !(c->unused_inode_hints = kcalloc(1U << c->inode_shard_bits, + sizeof(u64), GFP_KERNEL)) || bch2_io_clock_init(&c->io_clock[READ]) || bch2_io_clock_init(&c->io_clock[WRITE]) || bch2_fs_journal_init(&c->journal) || -- 2.30.2