X-Git-Url: https://git.sesse.net/?a=blobdiff_plain;ds=sidebyside;f=libbcachefs%2Fbtree_update_interior.c;h=9e6006d075851d8ad97a439ab3305fd4e154d943;hb=05408b6f8fea54bf53e68a4ef24291214970f6d0;hp=d84bb6806683ae8ba04efcd3532d4cdc9934946f;hpb=304691592738dc272f4150107b54a53ab43fc8be;p=bcachefs-tools-debian diff --git a/libbcachefs/btree_update_interior.c b/libbcachefs/btree_update_interior.c index d84bb68..9e6006d 100644 --- a/libbcachefs/btree_update_interior.c +++ b/libbcachefs/btree_update_interior.c @@ -21,50 +21,41 @@ #include #include -static void btree_node_will_make_reachable(struct btree_update *, - struct btree *); -static void btree_update_drop_new_node(struct bch_fs *, struct btree *); -static void bch2_btree_set_root_ondisk(struct bch_fs *, struct btree *, int); - /* Debug code: */ +/* + * Verify that child nodes correctly span parent node's range: + */ static void btree_node_interior_verify(struct btree *b) { +#ifdef CONFIG_BCACHEFS_DEBUG + struct bpos next_node = b->data->min_key; struct btree_node_iter iter; - struct bkey_packed *k; + struct bkey_s_c k; + struct bkey_s_c_btree_ptr_v2 bp; + struct bkey unpacked; - BUG_ON(!b->level); + BUG_ON(!b->c.level); - bch2_btree_node_iter_init(&iter, b, &b->key.k.p); -#if 1 - BUG_ON(!(k = bch2_btree_node_iter_peek(&iter, b)) || - bkey_cmp_left_packed(b, k, &b->key.k.p)); + bch2_btree_node_iter_init_from_start(&iter, b); - BUG_ON((bch2_btree_node_iter_advance(&iter, b), - !bch2_btree_node_iter_end(&iter))); -#else - const char *msg; + while (1) { + k = bch2_btree_node_iter_peek_unpack(&iter, b, &unpacked); + if (k.k->type != KEY_TYPE_btree_ptr_v2) + break; + bp = bkey_s_c_to_btree_ptr_v2(k); - msg = "not found"; - k = bch2_btree_node_iter_peek(&iter, b); - if (!k) - goto err; + BUG_ON(bkey_cmp(next_node, bp.v->min_key)); - msg = "isn't what it should be"; - if (bkey_cmp_left_packed(b, k, &b->key.k.p)) - goto err; + bch2_btree_node_iter_advance(&iter, b); - bch2_btree_node_iter_advance(&iter, b); + if (bch2_btree_node_iter_end(&iter)) { + BUG_ON(bkey_cmp(k.k->p, b->key.k.p)); + break; + } - msg = "isn't last key"; - if (!bch2_btree_node_iter_end(&iter)) - goto err; - return; -err: - bch2_dump_btree_node(b); - printk(KERN_ERR "last key %llu:%llu %s\n", b->key.k.p.inode, - b->key.k.p.offset, msg); - BUG(); + next_node = bkey_successor(k.k->p); + } #endif } @@ -129,74 +120,6 @@ bool bch2_btree_node_format_fits(struct bch_fs *c, struct btree *b, /* Btree node freeing/allocation: */ -static bool btree_key_matches(struct bch_fs *c, - struct bkey_s_c l, - struct bkey_s_c r) -{ - struct bkey_ptrs_c ptrs1 = bch2_bkey_ptrs_c(l); - struct bkey_ptrs_c ptrs2 = bch2_bkey_ptrs_c(r); - const struct bch_extent_ptr *ptr1, *ptr2; - - bkey_for_each_ptr(ptrs1, ptr1) - bkey_for_each_ptr(ptrs2, ptr2) - if (ptr1->dev == ptr2->dev && - ptr1->gen == ptr2->gen && - ptr1->offset == ptr2->offset) - return true; - - return false; -} - -/* - * We're doing the index update that makes @b unreachable, update stuff to - * reflect that: - * - * Must be called _before_ btree_update_updated_root() or - * btree_update_updated_node: - */ -static void bch2_btree_node_free_index(struct btree_update *as, struct btree *b, - struct bkey_s_c k, - struct bch_fs_usage *stats) -{ - struct bch_fs *c = as->c; - struct pending_btree_node_free *d; - - for (d = as->pending; d < as->pending + as->nr_pending; d++) - if (!bkey_cmp(k.k->p, d->key.k.p) && - btree_key_matches(c, k, bkey_i_to_s_c(&d->key))) - goto found; - BUG(); -found: - BUG_ON(d->index_update_done); - d->index_update_done = true; - - /* - * We're dropping @k from the btree, but it's still live until the - * index update is persistent so we need to keep a reference around for - * mark and sweep to find - that's primarily what the - * btree_node_pending_free list is for. - * - * So here (when we set index_update_done = true), we're moving an - * existing reference to a different part of the larger "gc keyspace" - - * and the new position comes after the old position, since GC marks - * the pending free list after it walks the btree. - * - * If we move the reference while mark and sweep is _between_ the old - * and the new position, mark and sweep will see the reference twice - * and it'll get double accounted - so check for that here and subtract - * to cancel out one of mark and sweep's markings if necessary: - */ - - if (gc_pos_cmp(c->gc_pos, b - ? gc_pos_btree_node(b) - : gc_pos_btree_root(as->btree_id)) >= 0 && - gc_pos_cmp(c->gc_pos, gc_phase(GC_PHASE_PENDING_DELETE)) < 0) - bch2_mark_key_locked(c, bkey_i_to_s_c(&d->key), - 0, 0, NULL, 0, - BTREE_TRIGGER_OVERWRITE| - BTREE_TRIGGER_GC); -} - static void __btree_node_free(struct bch_fs *c, struct btree *b) { trace_btree_node_free(c, b); @@ -212,6 +135,8 @@ static void __btree_node_free(struct bch_fs *c, struct btree *b) bch2_btree_node_hash_remove(&c->btree_cache, b); + six_lock_wakeup_all(&b->c.lock); + mutex_lock(&c->btree_cache.lock); list_move(&b->list, &c->btree_cache.freeable); mutex_unlock(&c->btree_cache.lock); @@ -221,15 +146,13 @@ void bch2_btree_node_free_never_inserted(struct bch_fs *c, struct btree *b) { struct open_buckets ob = b->ob; - btree_update_drop_new_node(c, b); - b->ob.nr = 0; clear_btree_node_dirty(b); btree_node_lock_type(c, b, SIX_LOCK_write); __btree_node_free(c, b); - six_unlock_write(&b->lock); + six_unlock_write(&b->c.lock); bch2_open_buckets_put(c, &ob); } @@ -240,38 +163,12 @@ void bch2_btree_node_free_inmem(struct bch_fs *c, struct btree *b, struct btree_iter *linked; trans_for_each_iter(iter->trans, linked) - BUG_ON(linked->l[b->level].b == b); - - /* - * Is this a node that isn't reachable on disk yet? - * - * Nodes that aren't reachable yet have writes blocked until they're - * reachable - now that we've cancelled any pending writes and moved - * things waiting on that write to wait on this update, we can drop this - * node from the list of nodes that the other update is making - * reachable, prior to freeing it: - */ - btree_update_drop_new_node(c, b); + BUG_ON(linked->l[b->c.level].b == b); - six_lock_write(&b->lock); + six_lock_write(&b->c.lock, NULL, NULL); __btree_node_free(c, b); - six_unlock_write(&b->lock); - six_unlock_intent(&b->lock); -} - -static void bch2_btree_node_free_ondisk(struct bch_fs *c, - struct pending_btree_node_free *pending) -{ - BUG_ON(!pending->index_update_done); - - bch2_mark_key(c, bkey_i_to_s_c(&pending->key), - 0, 0, NULL, 0, BTREE_TRIGGER_OVERWRITE); - - if (gc_visited(c, gc_phase(GC_PHASE_PENDING_DELETE))) - bch2_mark_key(c, bkey_i_to_s_c(&pending->key), - 0, 0, NULL, 0, - BTREE_TRIGGER_OVERWRITE| - BTREE_TRIGGER_GC); + six_unlock_write(&b->c.lock); + six_unlock_intent(&b->c.lock); } static struct btree *__bch2_btree_node_alloc(struct bch_fs *c, @@ -332,7 +229,11 @@ retry: goto retry; } - bkey_btree_ptr_init(&tmp.k); + if (c->sb.features & (1ULL << BCH_FEATURE_btree_ptr_v2)) + bkey_btree_ptr_v2_init(&tmp.k); + else + bkey_btree_ptr_init(&tmp.k); + bch2_alloc_sectors_append_ptrs(c, wp, &tmp.k, c->opts.btree_node_size); bch2_open_bucket_get(c, wp, &ob); @@ -354,25 +255,36 @@ static struct btree *bch2_btree_node_alloc(struct btree_update *as, unsigned lev { struct bch_fs *c = as->c; struct btree *b; + int ret; BUG_ON(level >= BTREE_MAX_DEPTH); - BUG_ON(!as->reserve->nr); - - b = as->reserve->b[--as->reserve->nr]; + BUG_ON(!as->nr_prealloc_nodes); - BUG_ON(bch2_btree_node_hash_insert(&c->btree_cache, b, level, as->btree_id)); + b = as->prealloc_nodes[--as->nr_prealloc_nodes]; set_btree_node_accessed(b); set_btree_node_dirty(b); set_btree_node_need_write(b); bch2_bset_init_first(b, &b->data->keys); + b->c.level = level; + b->c.btree_id = as->btree_id; + memset(&b->nr, 0, sizeof(b->nr)); b->data->magic = cpu_to_le64(bset_magic(c)); b->data->flags = 0; SET_BTREE_NODE_ID(b->data, as->btree_id); SET_BTREE_NODE_LEVEL(b->data, level); - b->data->ptr = bkey_i_to_btree_ptr(&b->key)->v.start[0]; + b->data->ptr = bch2_bkey_ptrs_c(bkey_i_to_s_c(&b->key)).start->ptr; + + if (b->key.k.type == KEY_TYPE_btree_ptr_v2) { + struct bkey_i_btree_ptr_v2 *bp = bkey_i_to_btree_ptr_v2(&b->key); + + bp->v.mem_ptr = 0; + bp->v.seq = b->data->keys.seq; + bp->v.sectors_written = 0; + bp->v.sectors = cpu_to_le16(c->opts.btree_node_size); + } if (c->sb.features & (1ULL << BCH_FEATURE_new_extent_overwrite)) SET_BTREE_NODE_NEW_EXTENT_OVERWRITE(b->data, true); @@ -383,25 +295,40 @@ static struct btree *bch2_btree_node_alloc(struct btree_update *as, unsigned lev bch2_btree_build_aux_trees(b); - btree_node_will_make_reachable(as, b); + ret = bch2_btree_node_hash_insert(&c->btree_cache, b, level, as->btree_id); + BUG_ON(ret); trace_btree_node_alloc(c, b); return b; } +static void btree_set_min(struct btree *b, struct bpos pos) +{ + if (b->key.k.type == KEY_TYPE_btree_ptr_v2) + bkey_i_to_btree_ptr_v2(&b->key)->v.min_key = pos; + b->data->min_key = pos; +} + +static void btree_set_max(struct btree *b, struct bpos pos) +{ + b->key.k.p = pos; + b->data->max_key = pos; +} + struct btree *__bch2_btree_node_alloc_replacement(struct btree_update *as, struct btree *b, struct bkey_format format) { struct btree *n; - n = bch2_btree_node_alloc(as, b->level); + n = bch2_btree_node_alloc(as, b->c.level); - n->data->min_key = b->data->min_key; - n->data->max_key = b->data->max_key; - n->data->format = format; SET_BTREE_NODE_SEQ(n->data, BTREE_NODE_SEQ(b->data) + 1); + btree_set_min(n, b->data->min_key); + btree_set_max(n, b->data->max_key); + + n->data->format = format; btree_node_set_format(n, format); bch2_btree_sort_into(as->c, n, b); @@ -431,29 +358,29 @@ static struct btree *__btree_root_alloc(struct btree_update *as, unsigned level) { struct btree *b = bch2_btree_node_alloc(as, level); - b->data->min_key = POS_MIN; - b->data->max_key = POS_MAX; + btree_set_min(b, POS_MIN); + btree_set_max(b, POS_MAX); b->data->format = bch2_btree_calc_format(b); - b->key.k.p = POS_MAX; btree_node_set_format(b, b->data->format); bch2_btree_build_aux_trees(b); - six_unlock_write(&b->lock); + bch2_btree_update_add_new_node(as, b); + six_unlock_write(&b->c.lock); return b; } -static void bch2_btree_reserve_put(struct bch_fs *c, struct btree_reserve *reserve) +static void bch2_btree_reserve_put(struct btree_update *as) { - bch2_disk_reservation_put(c, &reserve->disk_res); + struct bch_fs *c = as->c; mutex_lock(&c->btree_reserve_cache_lock); - while (reserve->nr) { - struct btree *b = reserve->b[--reserve->nr]; + while (as->nr_prealloc_nodes) { + struct btree *b = as->prealloc_nodes[--as->nr_prealloc_nodes]; - six_unlock_write(&b->lock); + six_unlock_write(&b->c.lock); if (c->btree_reserve_cache_nr < ARRAY_SIZE(c->btree_reserve_cache)) { @@ -469,42 +396,20 @@ static void bch2_btree_reserve_put(struct bch_fs *c, struct btree_reserve *reser btree_node_lock_type(c, b, SIX_LOCK_write); __btree_node_free(c, b); - six_unlock_write(&b->lock); + six_unlock_write(&b->c.lock); - six_unlock_intent(&b->lock); + six_unlock_intent(&b->c.lock); } mutex_unlock(&c->btree_reserve_cache_lock); - - mempool_free(reserve, &c->btree_reserve_pool); } -static struct btree_reserve *bch2_btree_reserve_get(struct bch_fs *c, - unsigned nr_nodes, - unsigned flags, - struct closure *cl) +static int bch2_btree_reserve_get(struct btree_update *as, unsigned nr_nodes, + unsigned flags, struct closure *cl) { - struct btree_reserve *reserve; + struct bch_fs *c = as->c; struct btree *b; - struct disk_reservation disk_res = { 0, 0 }; - unsigned sectors = nr_nodes * c->opts.btree_node_size; - int ret, disk_res_flags = 0; - - if (flags & BTREE_INSERT_NOFAIL) - disk_res_flags |= BCH_DISK_RESERVATION_NOFAIL; - - /* - * This check isn't necessary for correctness - it's just to potentially - * prevent us from doing a lot of work that'll end up being wasted: - */ - ret = bch2_journal_error(&c->journal); - if (ret) - return ERR_PTR(ret); - - if (bch2_disk_reservation_get(c, &disk_res, sectors, - c->opts.metadata_replicas, - disk_res_flags)) - return ERR_PTR(-ENOSPC); + int ret; BUG_ON(nr_nodes > BTREE_RESERVE_MAX); @@ -513,18 +418,11 @@ static struct btree_reserve *bch2_btree_reserve_get(struct bch_fs *c, * open bucket reserve: */ ret = bch2_btree_cache_cannibalize_lock(c, cl); - if (ret) { - bch2_disk_reservation_put(c, &disk_res); - return ERR_PTR(ret); - } - - reserve = mempool_alloc(&c->btree_reserve_pool, GFP_NOIO); - - reserve->disk_res = disk_res; - reserve->nr = 0; + if (ret) + return ret; - while (reserve->nr < nr_nodes) { - b = __bch2_btree_node_alloc(c, &disk_res, + while (as->nr_prealloc_nodes < nr_nodes) { + b = __bch2_btree_node_alloc(c, &as->disk_res, flags & BTREE_INSERT_NOWAIT ? NULL : cl, flags); if (IS_ERR(b)) { @@ -536,16 +434,15 @@ static struct btree_reserve *bch2_btree_reserve_get(struct bch_fs *c, if (ret) goto err_free; - reserve->b[reserve->nr++] = b; + as->prealloc_nodes[as->nr_prealloc_nodes++] = b; } bch2_btree_cache_cannibalize_unlock(c); - return reserve; + return 0; err_free: - bch2_btree_reserve_put(c, reserve); bch2_btree_cache_cannibalize_unlock(c); trace_btree_reserve_get_fail(c, nr_nodes, cl); - return ERR_PTR(ret); + return ret; } /* Asynchronous interior node update machinery */ @@ -554,191 +451,199 @@ static void bch2_btree_update_free(struct btree_update *as) { struct bch_fs *c = as->c; - bch2_journal_pin_flush(&c->journal, &as->journal); - - BUG_ON(as->nr_new_nodes); - BUG_ON(as->nr_pending); + bch2_journal_preres_put(&c->journal, &as->journal_preres); - if (as->reserve) - bch2_btree_reserve_put(c, as->reserve); + bch2_journal_pin_drop(&c->journal, &as->journal); + bch2_journal_pin_flush(&c->journal, &as->journal); + bch2_disk_reservation_put(c, &as->disk_res); + bch2_btree_reserve_put(as); mutex_lock(&c->btree_interior_update_lock); + list_del(&as->unwritten_list); list_del(&as->list); + mutex_unlock(&c->btree_interior_update_lock); closure_debug_destroy(&as->cl); mempool_free(as, &c->btree_interior_update_pool); closure_wake_up(&c->btree_interior_update_wait); - mutex_unlock(&c->btree_interior_update_lock); } -static void btree_update_nodes_reachable(struct closure *cl) +static void btree_update_will_delete_key(struct btree_update *as, + struct bkey_i *k) { - struct btree_update *as = container_of(cl, struct btree_update, cl); - struct bch_fs *c = as->c; - - bch2_journal_pin_drop(&c->journal, &as->journal); - - mutex_lock(&c->btree_interior_update_lock); - - while (as->nr_new_nodes) { - struct btree *b = as->new_nodes[--as->nr_new_nodes]; - - BUG_ON(b->will_make_reachable != (unsigned long) as); - b->will_make_reachable = 0; - mutex_unlock(&c->btree_interior_update_lock); - - /* - * b->will_make_reachable prevented it from being written, so - * write it now if it needs to be written: - */ - btree_node_lock_type(c, b, SIX_LOCK_read); - bch2_btree_node_write_cond(c, b, btree_node_need_write(b)); - six_unlock_read(&b->lock); - mutex_lock(&c->btree_interior_update_lock); - } - - while (as->nr_pending) - bch2_btree_node_free_ondisk(c, &as->pending[--as->nr_pending]); - - mutex_unlock(&c->btree_interior_update_lock); - - closure_wake_up(&as->wait); + BUG_ON(bch2_keylist_u64s(&as->old_keys) + k->k.u64s > + ARRAY_SIZE(as->_old_keys)); + bch2_keylist_add(&as->old_keys, k); +} - bch2_btree_update_free(as); +static void btree_update_will_add_key(struct btree_update *as, + struct bkey_i *k) +{ + BUG_ON(bch2_keylist_u64s(&as->new_keys) + k->k.u64s > + ARRAY_SIZE(as->_new_keys)); + bch2_keylist_add(&as->new_keys, k); } -static void btree_update_wait_on_journal(struct closure *cl) +/* + * The transactional part of an interior btree node update, where we journal the + * update we did to the interior node and update alloc info: + */ +static int btree_update_nodes_written_trans(struct btree_trans *trans, + struct btree_update *as) { - struct btree_update *as = container_of(cl, struct btree_update, cl); - struct bch_fs *c = as->c; + struct bkey_i *k; int ret; - ret = bch2_journal_open_seq_async(&c->journal, as->journal_seq, cl); - if (ret == -EAGAIN) { - continue_at(cl, btree_update_wait_on_journal, system_wq); - return; + trans->extra_journal_entries = (void *) &as->journal_entries[0]; + trans->extra_journal_entry_u64s = as->journal_u64s; + trans->journal_pin = &as->journal; + + for_each_keylist_key(&as->new_keys, k) { + ret = bch2_trans_mark_key(trans, bkey_i_to_s_c(k), + 0, 0, BTREE_TRIGGER_INSERT); + if (ret) + return ret; } - if (ret < 0) - goto err; - bch2_journal_flush_seq_async(&c->journal, as->journal_seq, cl); -err: - continue_at(cl, btree_update_nodes_reachable, system_wq); + for_each_keylist_key(&as->old_keys, k) { + ret = bch2_trans_mark_key(trans, bkey_i_to_s_c(k), + 0, 0, BTREE_TRIGGER_OVERWRITE); + if (ret) + return ret; + } + + return 0; } -static void btree_update_nodes_written(struct closure *cl) +static void btree_update_nodes_written(struct btree_update *as) { - struct btree_update *as = container_of(cl, struct btree_update, cl); struct bch_fs *c = as->c; - struct btree *b; + struct btree *b = as->b; + u64 journal_seq = 0; + unsigned i; + int ret; /* * We did an update to a parent node where the pointers we added pointed * to child nodes that weren't written yet: now, the child nodes have * been written so we can write out the update to the interior node. */ -retry: - mutex_lock(&c->btree_interior_update_lock); - as->nodes_written = true; - switch (as->mode) { - case BTREE_INTERIOR_NO_UPDATE: - BUG(); - case BTREE_INTERIOR_UPDATING_NODE: - /* The usual case: */ - b = READ_ONCE(as->b); - - if (!six_trylock_read(&b->lock)) { - mutex_unlock(&c->btree_interior_update_lock); - btree_node_lock_type(c, b, SIX_LOCK_read); - six_unlock_read(&b->lock); - goto retry; - } + /* + * We can't call into journal reclaim here: we'd block on the journal + * reclaim lock, but we may need to release the open buckets we have + * pinned in order for other btree updates to make forward progress, and + * journal reclaim does btree updates when flushing bkey_cached entries, + * which may require allocations as well. + */ + ret = bch2_trans_do(c, &as->disk_res, &journal_seq, + BTREE_INSERT_NOFAIL| + BTREE_INSERT_USE_RESERVE| + BTREE_INSERT_USE_ALLOC_RESERVE| + BTREE_INSERT_NOCHECK_RW| + BTREE_INSERT_JOURNAL_RECLAIM| + BTREE_INSERT_JOURNAL_RESERVED, + btree_update_nodes_written_trans(&trans, as)); + BUG_ON(ret && !bch2_journal_error(&c->journal)); + + if (b) { + /* + * @b is the node we did the final insert into: + * + * On failure to get a journal reservation, we still have to + * unblock the write and allow most of the write path to happen + * so that shutdown works, but the i->journal_seq mechanism + * won't work to prevent the btree write from being visible (we + * didn't get a journal sequence number) - instead + * __bch2_btree_node_write() doesn't do the actual write if + * we're in journal error state: + */ - BUG_ON(!btree_node_dirty(b)); - closure_wait(&btree_current_write(b)->wait, cl); + btree_node_lock_type(c, b, SIX_LOCK_intent); + btree_node_lock_type(c, b, SIX_LOCK_write); + mutex_lock(&c->btree_interior_update_lock); list_del(&as->write_blocked_list); - /* - * for flush_held_btree_writes() waiting on updates to flush or - * nodes to be writeable: - */ - closure_wake_up(&c->btree_interior_update_wait); - mutex_unlock(&c->btree_interior_update_lock); + if (!ret && as->b == b) { + struct bset *i = btree_bset_last(b); - /* - * b->write_blocked prevented it from being written, so - * write it now if it needs to be written: - */ - bch2_btree_node_write_cond(c, b, true); - six_unlock_read(&b->lock); - break; + BUG_ON(!b->c.level); + BUG_ON(!btree_node_dirty(b)); - case BTREE_INTERIOR_UPDATING_AS: - /* - * The btree node we originally updated has been freed and is - * being rewritten - so we need to write anything here, we just - * need to signal to that btree_update that it's ok to make the - * new replacement node visible: - */ - closure_put(&as->parent_as->cl); + i->journal_seq = cpu_to_le64( + max(journal_seq, + le64_to_cpu(i->journal_seq))); + + bch2_btree_add_journal_pin(c, b, journal_seq); + } - /* - * and then we have to wait on that btree_update to finish: - */ - closure_wait(&as->parent_as->wait, cl); mutex_unlock(&c->btree_interior_update_lock); - break; + six_unlock_write(&b->c.lock); - case BTREE_INTERIOR_UPDATING_ROOT: - /* b is the new btree root: */ - b = READ_ONCE(as->b); + btree_node_write_if_need(c, b, SIX_LOCK_intent); + six_unlock_intent(&b->c.lock); + } - if (!six_trylock_read(&b->lock)) { - mutex_unlock(&c->btree_interior_update_lock); - btree_node_lock_type(c, b, SIX_LOCK_read); - six_unlock_read(&b->lock); - goto retry; - } + bch2_journal_pin_drop(&c->journal, &as->journal); - BUG_ON(c->btree_roots[b->btree_id].as != as); - c->btree_roots[b->btree_id].as = NULL; + bch2_journal_preres_put(&c->journal, &as->journal_preres); - bch2_btree_set_root_ondisk(c, b, WRITE); + mutex_lock(&c->btree_interior_update_lock); + for (i = 0; i < as->nr_new_nodes; i++) { + b = as->new_nodes[i]; - /* - * We don't have to wait anything anything here (before - * btree_update_nodes_reachable frees the old nodes - * ondisk) - we've ensured that the very next journal write will - * have the pointer to the new root, and before the allocator - * can reuse the old nodes it'll have to do a journal commit: - */ - six_unlock_read(&b->lock); - mutex_unlock(&c->btree_interior_update_lock); + BUG_ON(b->will_make_reachable != (unsigned long) as); + b->will_make_reachable = 0; + } + mutex_unlock(&c->btree_interior_update_lock); - /* - * Bit of funny circularity going on here we have to break: - * - * We have to drop our journal pin before writing the journal - * entry that points to the new btree root: else, we could - * deadlock if the journal currently happens to be full. - * - * This mean we're dropping the journal pin _before_ the new - * nodes are technically reachable - but this is safe, because - * after the bch2_btree_set_root_ondisk() call above they will - * be reachable as of the very next journal write: - */ - bch2_journal_pin_drop(&c->journal, &as->journal); + for (i = 0; i < as->nr_new_nodes; i++) { + b = as->new_nodes[i]; - as->journal_seq = bch2_journal_last_unwritten_seq(&c->journal); + btree_node_lock_type(c, b, SIX_LOCK_read); + btree_node_write_if_need(c, b, SIX_LOCK_read); + six_unlock_read(&b->c.lock); + } - btree_update_wait_on_journal(cl); - return; + for (i = 0; i < as->nr_open_buckets; i++) + bch2_open_bucket_put(c, c->open_buckets + as->open_buckets[i]); + + bch2_btree_update_free(as); +} + +static void btree_interior_update_work(struct work_struct *work) +{ + struct bch_fs *c = + container_of(work, struct bch_fs, btree_interior_update_work); + struct btree_update *as; + + while (1) { + mutex_lock(&c->btree_interior_update_lock); + as = list_first_entry_or_null(&c->btree_interior_updates_unwritten, + struct btree_update, unwritten_list); + if (as && !as->nodes_written) + as = NULL; + mutex_unlock(&c->btree_interior_update_lock); + + if (!as) + break; + + btree_update_nodes_written(as); } +} + +static void btree_update_set_nodes_written(struct closure *cl) +{ + struct btree_update *as = container_of(cl, struct btree_update, cl); + struct bch_fs *c = as->c; + + mutex_lock(&c->btree_interior_update_lock); + as->nodes_written = true; + mutex_unlock(&c->btree_interior_update_lock); - continue_at(cl, btree_update_nodes_reachable, system_wq); + queue_work(c->btree_interior_update_worker, &c->btree_interior_update_work); } /* @@ -750,52 +655,16 @@ static void btree_update_updated_node(struct btree_update *as, struct btree *b) struct bch_fs *c = as->c; mutex_lock(&c->btree_interior_update_lock); + list_add_tail(&as->unwritten_list, &c->btree_interior_updates_unwritten); BUG_ON(as->mode != BTREE_INTERIOR_NO_UPDATE); BUG_ON(!btree_node_dirty(b)); - as->mode = BTREE_INTERIOR_UPDATING_NODE; - as->b = b; + as->mode = BTREE_INTERIOR_UPDATING_NODE; + as->b = b; list_add(&as->write_blocked_list, &b->write_blocked); mutex_unlock(&c->btree_interior_update_lock); - - /* - * In general, when you're staging things in a journal that will later - * be written elsewhere, and you also want to guarantee ordering: that - * is, if you have updates a, b, c, after a crash you should never see c - * and not a or b - there's a problem: - * - * If the final destination of the update(s) (i.e. btree node) can be - * written/flushed _before_ the relevant journal entry - oops, that - * breaks ordering, since the various leaf nodes can be written in any - * order. - * - * Normally we use bset->journal_seq to deal with this - if during - * recovery we find a btree node write that's newer than the newest - * journal entry, we just ignore it - we don't need it, anything we're - * supposed to have (that we reported as completed via fsync()) will - * still be in the journal, and as far as the state of the journal is - * concerned that btree node write never happened. - * - * That breaks when we're rewriting/splitting/merging nodes, since we're - * mixing btree node writes that haven't happened yet with previously - * written data that has been reported as completed to the journal. - * - * Thus, before making the new nodes reachable, we have to wait the - * newest journal sequence number we have data for to be written (if it - * hasn't been yet). - */ - bch2_journal_wait_on_seq(&c->journal, as->journal_seq, &as->cl); -} - -static void interior_update_flush(struct journal *j, - struct journal_entry_pin *pin, u64 seq) -{ - struct btree_update *as = - container_of(pin, struct btree_update, journal); - - bch2_journal_flush_seq_async(j, as->journal_seq, NULL); } static void btree_update_reparent(struct btree_update *as, @@ -803,10 +672,10 @@ static void btree_update_reparent(struct btree_update *as, { struct bch_fs *c = as->c; + lockdep_assert_held(&c->btree_interior_update_lock); + child->b = NULL; child->mode = BTREE_INTERIOR_UPDATING_AS; - child->parent_as = as; - closure_get(&as->cl); /* * When we write a new btree root, we have to drop our journal pin @@ -817,52 +686,51 @@ static void btree_update_reparent(struct btree_update *as, * just transfer the journal pin to the new interior update so * btree_update_nodes_written() can drop it. */ - bch2_journal_pin_add_if_older(&c->journal, &child->journal, - &as->journal, interior_update_flush); + bch2_journal_pin_copy(&c->journal, &as->journal, &child->journal, NULL); bch2_journal_pin_drop(&c->journal, &child->journal); - - as->journal_seq = max(as->journal_seq, child->journal_seq); } -static void btree_update_updated_root(struct btree_update *as) +static void btree_update_updated_root(struct btree_update *as, struct btree *b) { + struct bkey_i *insert = &b->key; struct bch_fs *c = as->c; - struct btree_root *r = &c->btree_roots[as->btree_id]; - - mutex_lock(&c->btree_interior_update_lock); BUG_ON(as->mode != BTREE_INTERIOR_NO_UPDATE); - /* - * Old root might not be persistent yet - if so, redirect its - * btree_update operation to point to us: - */ - if (r->as) - btree_update_reparent(as, r->as); + BUG_ON(as->journal_u64s + jset_u64s(insert->k.u64s) > + ARRAY_SIZE(as->journal_entries)); - as->mode = BTREE_INTERIOR_UPDATING_ROOT; - as->b = r->b; - r->as = as; + as->journal_u64s += + journal_entry_set((void *) &as->journal_entries[as->journal_u64s], + BCH_JSET_ENTRY_btree_root, + b->c.btree_id, b->c.level, + insert, insert->k.u64s); - mutex_unlock(&c->btree_interior_update_lock); + mutex_lock(&c->btree_interior_update_lock); + list_add_tail(&as->unwritten_list, &c->btree_interior_updates_unwritten); - /* - * When we're rewriting nodes and updating interior nodes, there's an - * issue with updates that haven't been written in the journal getting - * mixed together with older data - see btree_update_updated_node() - * for the explanation. - * - * However, this doesn't affect us when we're writing a new btree root - - * because to make that new root reachable we have to write out a new - * journal entry, which must necessarily be newer than as->journal_seq. - */ + as->mode = BTREE_INTERIOR_UPDATING_ROOT; + mutex_unlock(&c->btree_interior_update_lock); } -static void btree_node_will_make_reachable(struct btree_update *as, - struct btree *b) +/* + * bch2_btree_update_add_new_node: + * + * This causes @as to wait on @b to be written, before it gets to + * bch2_btree_update_nodes_written + * + * Additionally, it sets b->will_make_reachable to prevent any additional writes + * to @b from happening besides the first until @b is reachable on disk + * + * And it adds @b to the list of @as's new nodes, so that we can update sector + * counts in bch2_btree_update_nodes_written: + */ +void bch2_btree_update_add_new_node(struct btree_update *as, struct btree *b) { struct bch_fs *c = as->c; + closure_get(&as->cl); + mutex_lock(&c->btree_interior_update_lock); BUG_ON(as->nr_new_nodes >= ARRAY_SIZE(as->new_nodes)); BUG_ON(b->will_make_reachable); @@ -870,10 +738,14 @@ static void btree_node_will_make_reachable(struct btree_update *as, as->new_nodes[as->nr_new_nodes++] = b; b->will_make_reachable = 1UL|(unsigned long) as; - closure_get(&as->cl); mutex_unlock(&c->btree_interior_update_lock); + + btree_update_will_add_key(as, &b->key); } +/* + * returns true if @b was a new node + */ static void btree_update_drop_new_node(struct bch_fs *c, struct btree *b) { struct btree_update *as; @@ -881,6 +753,11 @@ static void btree_update_drop_new_node(struct bch_fs *c, struct btree *b) unsigned i; mutex_lock(&c->btree_interior_update_lock); + /* + * When b->will_make_reachable != 0, it owns a ref on as->cl that's + * dropped when it gets written by bch2_btree_complete_write - the + * xchg() is for synchronization with bch2_btree_complete_write: + */ v = xchg(&b->will_make_reachable, 0); as = (struct btree_update *) (v & ~1UL); @@ -902,25 +779,11 @@ found: closure_put(&as->cl); } -static void btree_interior_update_add_node_reference(struct btree_update *as, - struct btree *b) +void bch2_btree_update_get_open_buckets(struct btree_update *as, struct btree *b) { - struct bch_fs *c = as->c; - struct pending_btree_node_free *d; - - mutex_lock(&c->btree_interior_update_lock); - - /* Add this node to the list of nodes being freed: */ - BUG_ON(as->nr_pending >= ARRAY_SIZE(as->pending)); - - d = &as->pending[as->nr_pending++]; - d->index_update_done = false; - d->seq = b->data->keys.seq; - d->btree_id = b->btree_id; - d->level = b->level; - bkey_copy(&d->key, &b->key); - - mutex_unlock(&c->btree_interior_update_lock); + while (b->ob.nr) + as->open_buckets[as->nr_open_buckets++] = + b->ob.v[--b->ob.nr]; } /* @@ -932,30 +795,14 @@ void bch2_btree_interior_update_will_free_node(struct btree_update *as, struct btree *b) { struct bch_fs *c = as->c; - struct closure *cl, *cl_n; struct btree_update *p, *n; struct btree_write *w; - struct bset_tree *t; set_btree_node_dying(b); if (btree_node_fake(b)) return; - btree_interior_update_add_node_reference(as, b); - - /* - * Does this node have data that hasn't been written in the journal? - * - * If so, we have to wait for the corresponding journal entry to be - * written before making the new nodes reachable - we can't just carry - * over the bset->journal_seq tracking, since we'll be mixing those keys - * in with keys that aren't in the journal anymore: - */ - for_each_bset(b, t) - as->journal_seq = max(as->journal_seq, - le64_to_cpu(bset(b, t)->journal_seq)); - mutex_lock(&c->btree_interior_update_lock); /* @@ -967,7 +814,7 @@ void bch2_btree_interior_update_will_free_node(struct btree_update *as, * operations complete */ list_for_each_entry_safe(p, n, &b->write_blocked, write_blocked_list) { - list_del(&p->write_blocked_list); + list_del_init(&p->write_blocked_list); btree_update_reparent(as, p); /* @@ -979,16 +826,6 @@ void bch2_btree_interior_update_will_free_node(struct btree_update *as, clear_btree_node_dirty(b); clear_btree_node_need_write(b); - w = btree_current_write(b); - - /* - * Does this node have any btree_update operations waiting on this node - * to be written? - * - * If so, wake them up when this btree_update operation is reachable: - */ - llist_for_each_entry_safe(cl, cl_n, llist_del_all(&w->wait.list), list) - llist_add(&cl->list, &as->wait.list); /* * Does this node have unwritten data that has a pin on the journal? @@ -998,39 +835,59 @@ void bch2_btree_interior_update_will_free_node(struct btree_update *as, * oldest pin of any of the nodes we're freeing. We'll release the pin * when the new nodes are persistent and reachable on disk: */ - bch2_journal_pin_add_if_older(&c->journal, &w->journal, - &as->journal, interior_update_flush); + w = btree_current_write(b); + bch2_journal_pin_copy(&c->journal, &as->journal, &w->journal, NULL); bch2_journal_pin_drop(&c->journal, &w->journal); w = btree_prev_write(b); - bch2_journal_pin_add_if_older(&c->journal, &w->journal, - &as->journal, interior_update_flush); + bch2_journal_pin_copy(&c->journal, &as->journal, &w->journal, NULL); bch2_journal_pin_drop(&c->journal, &w->journal); mutex_unlock(&c->btree_interior_update_lock); + + /* + * Is this a node that isn't reachable on disk yet? + * + * Nodes that aren't reachable yet have writes blocked until they're + * reachable - now that we've cancelled any pending writes and moved + * things waiting on that write to wait on this update, we can drop this + * node from the list of nodes that the other update is making + * reachable, prior to freeing it: + */ + btree_update_drop_new_node(c, b); + + btree_update_will_delete_key(as, &b->key); } void bch2_btree_update_done(struct btree_update *as) { BUG_ON(as->mode == BTREE_INTERIOR_NO_UPDATE); - bch2_btree_reserve_put(as->c, as->reserve); - as->reserve = NULL; + bch2_btree_reserve_put(as); - continue_at(&as->cl, btree_update_nodes_written, system_freezable_wq); + continue_at(&as->cl, btree_update_set_nodes_written, system_freezable_wq); } struct btree_update * -bch2_btree_update_start(struct bch_fs *c, enum btree_id id, +bch2_btree_update_start(struct btree_trans *trans, enum btree_id id, unsigned nr_nodes, unsigned flags, struct closure *cl) { - struct btree_reserve *reserve; + struct bch_fs *c = trans->c; struct btree_update *as; + int disk_res_flags = (flags & BTREE_INSERT_NOFAIL) + ? BCH_DISK_RESERVATION_NOFAIL : 0; + int journal_flags = (flags & BTREE_INSERT_JOURNAL_RESERVED) + ? JOURNAL_RES_GET_RECLAIM : 0; + int ret = 0; - reserve = bch2_btree_reserve_get(c, nr_nodes, flags, cl); - if (IS_ERR(reserve)) - return ERR_CAST(reserve); + /* + * This check isn't necessary for correctness - it's just to potentially + * prevent us from doing a lot of work that'll end up being wasted: + */ + ret = bch2_journal_error(&c->journal); + if (ret) + return ERR_PTR(ret); as = mempool_alloc(&c->btree_interior_update_pool, GFP_NOIO); memset(as, 0, sizeof(*as)); @@ -1038,21 +895,58 @@ bch2_btree_update_start(struct bch_fs *c, enum btree_id id, as->c = c; as->mode = BTREE_INTERIOR_NO_UPDATE; as->btree_id = id; - as->reserve = reserve; + INIT_LIST_HEAD(&as->list); + INIT_LIST_HEAD(&as->unwritten_list); INIT_LIST_HEAD(&as->write_blocked_list); - + bch2_keylist_init(&as->old_keys, as->_old_keys); + bch2_keylist_init(&as->new_keys, as->_new_keys); bch2_keylist_init(&as->parent_keys, as->inline_keys); + ret = bch2_journal_preres_get(&c->journal, &as->journal_preres, + BTREE_UPDATE_JOURNAL_RES, + journal_flags|JOURNAL_RES_GET_NONBLOCK); + if (ret == -EAGAIN) { + if (flags & BTREE_INSERT_NOUNLOCK) + return ERR_PTR(-EINTR); + + bch2_trans_unlock(trans); + + ret = bch2_journal_preres_get(&c->journal, &as->journal_preres, + BTREE_UPDATE_JOURNAL_RES, + journal_flags); + if (ret) + return ERR_PTR(ret); + + if (!bch2_trans_relock(trans)) { + ret = -EINTR; + goto err; + } + } + + ret = bch2_disk_reservation_get(c, &as->disk_res, + nr_nodes * c->opts.btree_node_size, + c->opts.metadata_replicas, + disk_res_flags); + if (ret) + goto err; + + ret = bch2_btree_reserve_get(as, nr_nodes, flags, cl); + if (ret) + goto err; + mutex_lock(&c->btree_interior_update_lock); list_add_tail(&as->list, &c->btree_interior_update_list); mutex_unlock(&c->btree_interior_update_lock); return as; +err: + bch2_btree_update_free(as); + return ERR_PTR(ret); } /* Btree root updates: */ -static void __bch2_btree_set_root_inmem(struct bch_fs *c, struct btree *b) +static void bch2_btree_set_root_inmem(struct bch_fs *c, struct btree *b) { /* Root nodes cannot be reaped */ mutex_lock(&c->btree_cache.lock); @@ -1061,7 +955,7 @@ static void __bch2_btree_set_root_inmem(struct bch_fs *c, struct btree *b) mutex_lock(&c->btree_root_lock); BUG_ON(btree_node_root(c, b) && - (b->level < btree_node_root(c, b)->level || + (b->c.level < btree_node_root(c, b)->c.level || !btree_node_dying(btree_node_root(c, b)))); btree_node_root(c, b) = b; @@ -1070,54 +964,6 @@ static void __bch2_btree_set_root_inmem(struct bch_fs *c, struct btree *b) bch2_recalc_btree_reserve(c); } -static void bch2_btree_set_root_inmem(struct btree_update *as, struct btree *b) -{ - struct bch_fs *c = as->c; - struct btree *old = btree_node_root(c, b); - struct bch_fs_usage *fs_usage; - - __bch2_btree_set_root_inmem(c, b); - - mutex_lock(&c->btree_interior_update_lock); - percpu_down_read(&c->mark_lock); - fs_usage = bch2_fs_usage_scratch_get(c); - - bch2_mark_key_locked(c, bkey_i_to_s_c(&b->key), - 0, 0, fs_usage, 0, - BTREE_TRIGGER_INSERT); - if (gc_visited(c, gc_pos_btree_root(b->btree_id))) - bch2_mark_key_locked(c, bkey_i_to_s_c(&b->key), - 0, 0, NULL, 0, - BTREE_TRIGGER_INSERT| - BTREE_TRIGGER_GC); - - if (old && !btree_node_fake(old)) - bch2_btree_node_free_index(as, NULL, - bkey_i_to_s_c(&old->key), - fs_usage); - bch2_fs_usage_apply(c, fs_usage, &as->reserve->disk_res, 0); - - bch2_fs_usage_scratch_put(c, fs_usage); - percpu_up_read(&c->mark_lock); - mutex_unlock(&c->btree_interior_update_lock); -} - -static void bch2_btree_set_root_ondisk(struct bch_fs *c, struct btree *b, int rw) -{ - struct btree_root *r = &c->btree_roots[b->btree_id]; - - mutex_lock(&c->btree_root_lock); - - BUG_ON(b != r->b); - bkey_copy(&r->key, &b->key); - r->level = b->level; - r->alive = true; - if (rw == WRITE) - c->btree_roots_dirty = true; - - mutex_unlock(&c->btree_root_lock); -} - /** * bch_btree_set_root - update the root in memory and on disk * @@ -1148,9 +994,9 @@ static void bch2_btree_set_root(struct btree_update *as, struct btree *b, */ bch2_btree_node_lock_write(old, iter); - bch2_btree_set_root_inmem(as, b); + bch2_btree_set_root_inmem(c, b); - btree_update_updated_root(as); + btree_update_updated_root(as, b); /* * Unlock old root after new root is visible: @@ -1169,46 +1015,21 @@ static void bch2_insert_fixup_btree_ptr(struct btree_update *as, struct btree *b struct bkey_i *insert, struct btree_node_iter *node_iter) { - struct bch_fs *c = as->c; - struct bch_fs_usage *fs_usage; struct bkey_packed *k; - struct bkey tmp; - - BUG_ON(insert->k.u64s > bch_btree_keys_u64s_remaining(c, b)); - - mutex_lock(&c->btree_interior_update_lock); - percpu_down_read(&c->mark_lock); - fs_usage = bch2_fs_usage_scratch_get(c); - bch2_mark_key_locked(c, bkey_i_to_s_c(insert), - 0, 0, fs_usage, 0, - BTREE_TRIGGER_INSERT); + BUG_ON(as->journal_u64s + jset_u64s(insert->k.u64s) > + ARRAY_SIZE(as->journal_entries)); - if (gc_visited(c, gc_pos_btree_node(b))) - bch2_mark_key_locked(c, bkey_i_to_s_c(insert), - 0, 0, NULL, 0, - BTREE_TRIGGER_INSERT| - BTREE_TRIGGER_GC); + as->journal_u64s += + journal_entry_set((void *) &as->journal_entries[as->journal_u64s], + BCH_JSET_ENTRY_btree_keys, + b->c.btree_id, b->c.level, + insert, insert->k.u64s); while ((k = bch2_btree_node_iter_peek_all(node_iter, b)) && - bkey_iter_pos_cmp(b, &insert->k.p, k) > 0) + bkey_iter_pos_cmp(b, k, &insert->k.p) < 0) bch2_btree_node_iter_advance(node_iter, b); - /* - * If we're overwriting, look up pending delete and mark so that gc - * marks it on the pending delete list: - */ - if (k && !bkey_cmp_packed(b, k, &insert->k)) - bch2_btree_node_free_index(as, b, - bkey_disassemble(b, k, &tmp), - fs_usage); - - bch2_fs_usage_apply(c, fs_usage, &as->reserve->disk_res, 0); - - bch2_fs_usage_scratch_put(c, fs_usage); - percpu_up_read(&c->mark_lock); - mutex_unlock(&c->btree_interior_update_lock); - bch2_btree_bset_insert_key(iter, b, node_iter, insert); set_btree_node_dirty(b); set_btree_node_need_write(b); @@ -1227,7 +1048,8 @@ static struct btree *__btree_split_node(struct btree_update *as, struct bset *set1, *set2; struct bkey_packed *k, *prev = NULL; - n2 = bch2_btree_node_alloc(as, n1->level); + n2 = bch2_btree_node_alloc(as, n1->c.level); + bch2_btree_update_add_new_node(as, n2); n2->data->max_key = n1->data->max_key; n2->data->format = n1->format; @@ -1263,10 +1085,8 @@ static struct btree *__btree_split_node(struct btree_update *as, BUG_ON(!prev); - n1->key.k.p = bkey_unpack_pos(n1, prev); - n1->data->max_key = n1->key.k.p; - n2->data->min_key = - btree_type_successor(n1->btree_id, n1->key.k.p); + btree_set_max(n1, bkey_unpack_pos(n1, prev)); + btree_set_min(n2, bkey_successor(n1->key.k.p)); set2->u64s = cpu_to_le16((u64 *) vstruct_end(set1) - (u64 *) k); set1->u64s = cpu_to_le16(le16_to_cpu(set1->u64s) - le16_to_cpu(set2->u64s)); @@ -1297,7 +1117,7 @@ static struct btree *__btree_split_node(struct btree_update *as, bch2_verify_btree_nr_keys(n1); bch2_verify_btree_nr_keys(n2); - if (n1->level) { + if (n1->c.level) { btree_node_interior_verify(n1); btree_node_interior_verify(n2); } @@ -1332,11 +1152,6 @@ static void btree_split_insert_keys(struct btree_update *as, struct btree *b, while (!bch2_keylist_empty(keys)) { k = bch2_keylist_front(keys); - BUG_ON(bch_keylist_u64s(keys) > - bch_btree_keys_u64s_remaining(as->c, b)); - BUG_ON(bkey_cmp(k->k.p, b->data->min_key) < 0); - BUG_ON(bkey_cmp(k->k.p, b->data->max_key) > 0); - bch2_insert_fixup_btree_ptr(as, b, iter, k, &node_iter); bch2_keylist_pop_front(keys); } @@ -1376,24 +1191,25 @@ static void btree_split(struct btree_update *as, struct btree *b, u64 start_time = local_clock(); BUG_ON(!parent && (b != btree_node_root(c, b))); - BUG_ON(!btree_node_intent_locked(iter, btree_node_root(c, b)->level)); + BUG_ON(!btree_node_intent_locked(iter, btree_node_root(c, b)->c.level)); bch2_btree_interior_update_will_free_node(as, b); n1 = bch2_btree_node_alloc_replacement(as, b); + bch2_btree_update_add_new_node(as, n1); if (keys) btree_split_insert_keys(as, n1, iter, keys); - if (vstruct_blocks(n1->data, c->block_bits) > BTREE_SPLIT_THRESHOLD(c)) { + if (bset_u64s(&n1->set[0]) > BTREE_SPLIT_THRESHOLD(c)) { trace_btree_split(c, b); n2 = __btree_split_node(as, n1, iter); bch2_btree_build_aux_trees(n2); bch2_btree_build_aux_trees(n1); - six_unlock_write(&n2->lock); - six_unlock_write(&n1->lock); + six_unlock_write(&n2->c.lock); + six_unlock_write(&n1->c.lock); bch2_btree_node_write(c, n2, SIX_LOCK_intent); @@ -1407,7 +1223,7 @@ static void btree_split(struct btree_update *as, struct btree *b, if (!parent) { /* Depth increases, make a new root */ - n3 = __btree_root_alloc(as, b->level + 1); + n3 = __btree_root_alloc(as, b->c.level + 1); n3->sib_u64s[0] = U16_MAX; n3->sib_u64s[1] = U16_MAX; @@ -1420,9 +1236,10 @@ static void btree_split(struct btree_update *as, struct btree *b, trace_btree_compact(c, b); bch2_btree_build_aux_trees(n1); - six_unlock_write(&n1->lock); + six_unlock_write(&n1->c.lock); - bch2_keylist_add(&as->parent_keys, &n1->key); + if (parent) + bch2_keylist_add(&as->parent_keys, &n1->key); } bch2_btree_node_write(c, n1, SIX_LOCK_intent); @@ -1439,15 +1256,15 @@ static void btree_split(struct btree_update *as, struct btree *b, bch2_btree_set_root(as, n1, iter); } - bch2_open_buckets_put(c, &n1->ob); + bch2_btree_update_get_open_buckets(as, n1); if (n2) - bch2_open_buckets_put(c, &n2->ob); + bch2_btree_update_get_open_buckets(as, n2); if (n3) - bch2_open_buckets_put(c, &n3->ob); + bch2_btree_update_get_open_buckets(as, n3); /* Successful split, update the iterator to point to the new nodes: */ - six_lock_increment(&b->lock, SIX_LOCK_intent); + six_lock_increment(&b->c.lock, SIX_LOCK_intent); bch2_btree_iter_node_drop(iter, b); if (n3) bch2_btree_iter_node_replace(iter, n3); @@ -1464,10 +1281,10 @@ static void btree_split(struct btree_update *as, struct btree *b, bch2_btree_node_free_inmem(c, b, iter); if (n3) - six_unlock_intent(&n3->lock); + six_unlock_intent(&n3->c.lock); if (n2) - six_unlock_intent(&n2->lock); - six_unlock_intent(&n1->lock); + six_unlock_intent(&n2->c.lock); + six_unlock_intent(&n1->c.lock); bch2_btree_trans_verify_locks(iter->trans); @@ -1485,7 +1302,7 @@ bch2_btree_insert_keys_interior(struct btree_update *as, struct btree *b, struct bkey_packed *k; /* Don't screw up @iter's position: */ - node_iter = iter->l[b->level].iter; + node_iter = iter->l[b->c.level].iter; /* * btree_split(), btree_gc_coalesce() will insert keys before @@ -1496,19 +1313,15 @@ bch2_btree_insert_keys_interior(struct btree_update *as, struct btree *b, (bkey_cmp_packed(b, k, &insert->k) >= 0)) ; - while (!bch2_keylist_empty(keys)) { - insert = bch2_keylist_front(keys); - + for_each_keylist_key(keys, insert) bch2_insert_fixup_btree_ptr(as, b, iter, insert, &node_iter); - bch2_keylist_pop_front(keys); - } btree_update_updated_node(as, b); trans_for_each_iter_with_node(iter->trans, b, linked) - bch2_btree_node_iter_peek(&linked->l[b->level].iter, b); + bch2_btree_node_iter_peek(&linked->l[b->c.level].iter, b); - bch2_btree_iter_verify(iter, b); + bch2_btree_trans_verify_iters(iter->trans, b); } /** @@ -1532,8 +1345,8 @@ void bch2_btree_insert_node(struct btree_update *as, struct btree *b, int old_live_u64s = b->nr.live_u64s; int live_u64s_added, u64s_added; - BUG_ON(!btree_node_intent_locked(iter, btree_node_root(c, b)->level)); - BUG_ON(!b->level); + BUG_ON(!btree_node_intent_locked(iter, btree_node_root(c, b)->c.level)); + BUG_ON(!b->c.level); BUG_ON(!as || as->b); bch2_verify_keylist_sorted(keys); @@ -1542,7 +1355,7 @@ void bch2_btree_insert_node(struct btree_update *as, struct btree *b, bch2_btree_node_lock_for_insert(c, b, iter); - if (!bch2_btree_node_insert_fits(c, b, bch_keylist_u64s(keys))) { + if (!bch2_btree_node_insert_fits(c, b, bch2_keylist_u64s(keys))) { bch2_btree_node_unlock_write(b, iter); goto split; } @@ -1570,7 +1383,7 @@ void bch2_btree_insert_node(struct btree_update *as, struct btree *b, * the btree iterator yet, so the merge path's unlock/wait/relock dance * won't work: */ - bch2_foreground_maybe_merge(c, iter, b->level, + bch2_foreground_maybe_merge(c, iter, b->c.level, flags|BTREE_INSERT_NOUNLOCK); return; split: @@ -1581,7 +1394,7 @@ int bch2_btree_split_leaf(struct bch_fs *c, struct btree_iter *iter, unsigned flags) { struct btree_trans *trans = iter->trans; - struct btree *b = iter->l[0].b; + struct btree *b = iter_l(iter)->b; struct btree_update *as; struct closure cl; int ret = 0; @@ -1600,8 +1413,10 @@ int bch2_btree_split_leaf(struct bch_fs *c, struct btree_iter *iter, /* Hack, because gc and splitting nodes doesn't mix yet: */ if (!(flags & BTREE_INSERT_GC_LOCK_HELD) && !down_read_trylock(&c->gc_lock)) { - if (flags & BTREE_INSERT_NOUNLOCK) + if (flags & BTREE_INSERT_NOUNLOCK) { + trace_transaction_restart_ip(trans->ip, _THIS_IP_); return -EINTR; + } bch2_trans_unlock(trans); down_read(&c->gc_lock); @@ -1620,7 +1435,7 @@ int bch2_btree_split_leaf(struct bch_fs *c, struct btree_iter *iter, goto out; } - as = bch2_btree_update_start(c, iter->btree_id, + as = bch2_btree_update_start(trans, iter->btree_id, btree_update_reserve_required(c, b), flags, !(flags & BTREE_INSERT_NOUNLOCK) ? &cl : NULL); if (IS_ERR(as)) { @@ -1629,6 +1444,8 @@ int bch2_btree_split_leaf(struct bch_fs *c, struct btree_iter *iter, BUG_ON(flags & BTREE_INSERT_NOUNLOCK); bch2_trans_unlock(trans); ret = -EINTR; + + trace_transaction_restart_ip(trans->ip, _THIS_IP_); } goto out; } @@ -1718,7 +1535,7 @@ retry: b->sib_u64s[sib] = sib_u64s; if (b->sib_u64s[sib] > BTREE_FOREGROUND_MERGE_THRESHOLD(c)) { - six_unlock_intent(&m->lock); + six_unlock_intent(&m->c.lock); goto out; } @@ -1732,8 +1549,9 @@ retry: goto err_unlock; } - as = bch2_btree_update_start(c, iter->btree_id, + as = bch2_btree_update_start(trans, iter->btree_id, btree_update_reserve_required(c, parent) + 1, + flags| BTREE_INSERT_NOFAIL| BTREE_INSERT_USE_RESERVE, !(flags & BTREE_INSERT_NOUNLOCK) ? &cl : NULL); @@ -1747,12 +1565,12 @@ retry: bch2_btree_interior_update_will_free_node(as, b); bch2_btree_interior_update_will_free_node(as, m); - n = bch2_btree_node_alloc(as, b->level); + n = bch2_btree_node_alloc(as, b->c.level); + bch2_btree_update_add_new_node(as, n); - n->data->min_key = prev->data->min_key; - n->data->max_key = next->data->max_key; + btree_set_min(n, prev->data->min_key); + btree_set_max(n, next->data->max_key); n->data->format = new_f; - n->key.k.p = next->key.k.p; btree_node_set_format(n, new_f); @@ -1760,7 +1578,7 @@ retry: bch2_btree_sort_into(c, n, next); bch2_btree_build_aux_trees(n); - six_unlock_write(&n->lock); + six_unlock_write(&n->c.lock); bkey_init(&delete.k); delete.k.p = prev->key.k.p; @@ -1771,20 +1589,20 @@ retry: bch2_btree_insert_node(as, parent, iter, &as->parent_keys, flags); - bch2_open_buckets_put(c, &n->ob); + bch2_btree_update_get_open_buckets(as, n); - six_lock_increment(&b->lock, SIX_LOCK_intent); + six_lock_increment(&b->c.lock, SIX_LOCK_intent); bch2_btree_iter_node_drop(iter, b); bch2_btree_iter_node_drop(iter, m); bch2_btree_iter_node_replace(iter, n); - bch2_btree_iter_verify(iter, n); + bch2_btree_trans_verify_iters(trans, n); bch2_btree_node_free_inmem(c, b, iter); bch2_btree_node_free_inmem(c, m, iter); - six_unlock_intent(&n->lock); + six_unlock_intent(&n->c.lock); bch2_btree_update_done(as); @@ -1806,7 +1624,7 @@ out: return; err_cycle_gc_lock: - six_unlock_intent(&m->lock); + six_unlock_intent(&m->c.lock); if (flags & BTREE_INSERT_NOUNLOCK) goto out; @@ -1819,7 +1637,7 @@ err_cycle_gc_lock: goto err; err_unlock: - six_unlock_intent(&m->lock); + six_unlock_intent(&m->c.lock); if (!(flags & BTREE_INSERT_GC_LOCK_HELD)) up_read(&c->gc_lock); err: @@ -1846,7 +1664,7 @@ static int __btree_node_rewrite(struct bch_fs *c, struct btree_iter *iter, struct btree *n, *parent = btree_node_parent(iter, b); struct btree_update *as; - as = bch2_btree_update_start(c, iter->btree_id, + as = bch2_btree_update_start(iter->trans, iter->btree_id, (parent ? btree_update_reserve_required(c, parent) : 0) + 1, @@ -1859,9 +1677,10 @@ static int __btree_node_rewrite(struct bch_fs *c, struct btree_iter *iter, bch2_btree_interior_update_will_free_node(as, b); n = bch2_btree_node_alloc_replacement(as, b); + bch2_btree_update_add_new_node(as, n); bch2_btree_build_aux_trees(n); - six_unlock_write(&n->lock); + six_unlock_write(&n->c.lock); trace_btree_gc_rewrite_node(c, b); @@ -1874,13 +1693,13 @@ static int __btree_node_rewrite(struct bch_fs *c, struct btree_iter *iter, bch2_btree_set_root(as, n, iter); } - bch2_open_buckets_put(c, &n->ob); + bch2_btree_update_get_open_buckets(as, n); - six_lock_increment(&b->lock, SIX_LOCK_intent); + six_lock_increment(&b->c.lock, SIX_LOCK_intent); bch2_btree_iter_node_drop(iter, b); bch2_btree_iter_node_replace(iter, n); bch2_btree_node_free_inmem(c, b, iter); - six_unlock_intent(&n->lock); + six_unlock_intent(&n->c.lock); bch2_btree_update_done(as); return 0; @@ -1944,65 +1763,24 @@ static void __bch2_btree_node_update_key(struct bch_fs *c, struct btree_update *as, struct btree_iter *iter, struct btree *b, struct btree *new_hash, - struct bkey_i_btree_ptr *new_key) + struct bkey_i *new_key) { struct btree *parent; int ret; - /* - * Two corner cases that need to be thought about here: - * - * @b may not be reachable yet - there might be another interior update - * operation waiting on @b to be written, and we're gonna deliver the - * write completion to that interior update operation _before_ - * persisting the new_key update - * - * That ends up working without us having to do anything special here: - * the reason is, we do kick off (and do the in memory updates) for the - * update for @new_key before we return, creating a new interior_update - * operation here. - * - * The new interior update operation here will in effect override the - * previous one. The previous one was going to terminate - make @b - * reachable - in one of two ways: - * - updating the btree root pointer - * In that case, - * no, this doesn't work. argh. - */ - - if (b->will_make_reachable) - as->must_rewrite = true; - - btree_interior_update_add_node_reference(as, b); - - /* - * XXX: the rest of the update path treats this like we're actually - * inserting a new node and deleting the existing node, so the - * reservation needs to include enough space for @b - * - * that is actually sketch as fuck though and I am surprised the code - * seems to work like that, definitely need to go back and rework it - * into something saner. - * - * (I think @b is just getting double counted until the btree update - * finishes and "deletes" @b on disk) - */ - ret = bch2_disk_reservation_add(c, &as->reserve->disk_res, - c->opts.btree_node_size * - bch2_bkey_nr_ptrs(bkey_i_to_s_c(&new_key->k_i)), - BCH_DISK_RESERVATION_NOFAIL); - BUG_ON(ret); + btree_update_will_delete_key(as, &b->key); + btree_update_will_add_key(as, new_key); parent = btree_node_parent(iter, b); if (parent) { if (new_hash) { - bkey_copy(&new_hash->key, &new_key->k_i); + bkey_copy(&new_hash->key, new_key); ret = bch2_btree_node_hash_insert(&c->btree_cache, - new_hash, b->level, b->btree_id); + new_hash, b->c.level, b->c.btree_id); BUG_ON(ret); } - bch2_keylist_add(&as->parent_keys, &new_key->k_i); + bch2_keylist_add(&as->parent_keys, new_key); bch2_btree_insert_node(as, parent, iter, &as->parent_keys, 0); if (new_hash) { @@ -2011,55 +1789,29 @@ static void __bch2_btree_node_update_key(struct bch_fs *c, bch2_btree_node_hash_remove(&c->btree_cache, b); - bkey_copy(&b->key, &new_key->k_i); + bkey_copy(&b->key, new_key); ret = __bch2_btree_node_hash_insert(&c->btree_cache, b); BUG_ON(ret); mutex_unlock(&c->btree_cache.lock); } else { - bkey_copy(&b->key, &new_key->k_i); + bkey_copy(&b->key, new_key); } } else { - struct bch_fs_usage *fs_usage; - BUG_ON(btree_node_root(c, b) != b); bch2_btree_node_lock_write(b, iter); + bkey_copy(&b->key, new_key); - mutex_lock(&c->btree_interior_update_lock); - percpu_down_read(&c->mark_lock); - fs_usage = bch2_fs_usage_scratch_get(c); - - bch2_mark_key_locked(c, bkey_i_to_s_c(&new_key->k_i), - 0, 0, fs_usage, 0, - BTREE_TRIGGER_INSERT); - if (gc_visited(c, gc_pos_btree_root(b->btree_id))) - bch2_mark_key_locked(c, bkey_i_to_s_c(&new_key->k_i), - 0, 0, NULL, 0, - BTREE_TRIGGER_INSERT|| - BTREE_TRIGGER_GC); - - bch2_btree_node_free_index(as, NULL, - bkey_i_to_s_c(&b->key), - fs_usage); - bch2_fs_usage_apply(c, fs_usage, &as->reserve->disk_res, 0); - - bch2_fs_usage_scratch_put(c, fs_usage); - percpu_up_read(&c->mark_lock); - mutex_unlock(&c->btree_interior_update_lock); - - if (PTR_HASH(&new_key->k_i) != PTR_HASH(&b->key)) { + if (btree_ptr_hash_val(&b->key) != b->hash_val) { mutex_lock(&c->btree_cache.lock); bch2_btree_node_hash_remove(&c->btree_cache, b); - bkey_copy(&b->key, &new_key->k_i); ret = __bch2_btree_node_hash_insert(&c->btree_cache, b); BUG_ON(ret); mutex_unlock(&c->btree_cache.lock); - } else { - bkey_copy(&b->key, &new_key->k_i); } - btree_update_updated_root(as); + btree_update_updated_root(as, b); bch2_btree_node_unlock_write(b, iter); } @@ -2068,7 +1820,7 @@ static void __bch2_btree_node_update_key(struct bch_fs *c, int bch2_btree_node_update_key(struct bch_fs *c, struct btree_iter *iter, struct btree *b, - struct bkey_i_btree_ptr *new_key) + struct bkey_i *new_key) { struct btree *parent = btree_node_parent(iter, b); struct btree_update *as = NULL; @@ -2091,8 +1843,11 @@ int bch2_btree_node_update_key(struct bch_fs *c, struct btree_iter *iter, } } - /* check PTR_HASH() after @b is locked by btree_iter_traverse(): */ - if (PTR_HASH(&new_key->k_i) != PTR_HASH(&b->key)) { + /* + * check btree_ptr_hash_val() after @b is locked by + * btree_iter_traverse(): + */ + if (btree_ptr_hash_val(new_key) != b->hash_val) { /* bch2_btree_reserve_get will unlock */ ret = bch2_btree_cache_cannibalize_lock(c, &cl); if (ret) { @@ -2110,7 +1865,7 @@ int bch2_btree_node_update_key(struct bch_fs *c, struct btree_iter *iter, new_hash = bch2_btree_node_mem_alloc(c); } - as = bch2_btree_update_start(c, iter->btree_id, + as = bch2_btree_update_start(iter->trans, iter->btree_id, parent ? btree_update_reserve_required(c, parent) : 0, BTREE_INSERT_NOFAIL| BTREE_INSERT_USE_RESERVE| @@ -2134,7 +1889,7 @@ int bch2_btree_node_update_key(struct bch_fs *c, struct btree_iter *iter, goto err; } - ret = bch2_mark_bkey_replicas(c, bkey_i_to_s_c(&new_key->k_i)); + ret = bch2_mark_bkey_replicas(c, bkey_i_to_s_c(new_key)); if (ret) goto err_free_update; @@ -2147,8 +1902,8 @@ err: list_move(&new_hash->list, &c->btree_cache.freeable); mutex_unlock(&c->btree_cache.lock); - six_unlock_write(&new_hash->lock); - six_unlock_intent(&new_hash->lock); + six_unlock_write(&new_hash->c.lock); + six_unlock_intent(&new_hash->c.lock); } up_read(&c->gc_lock); closure_sync(&cl); @@ -2168,7 +1923,7 @@ void bch2_btree_set_root_for_read(struct bch_fs *c, struct btree *b) { BUG_ON(btree_node_root(c, b)); - __bch2_btree_set_root_inmem(c, b); + bch2_btree_set_root_inmem(c, b); } void bch2_btree_root_alloc(struct bch_fs *c, enum btree_id id) @@ -2188,29 +1943,30 @@ void bch2_btree_root_alloc(struct bch_fs *c, enum btree_id id) bch2_btree_cache_cannibalize_unlock(c); set_btree_node_fake(b); - b->level = 0; - b->btree_id = id; + b->c.level = 0; + b->c.btree_id = id; bkey_btree_ptr_init(&b->key); b->key.k.p = POS_MAX; - PTR_HASH(&b->key) = U64_MAX - id; + *((u64 *) bkey_i_to_btree_ptr(&b->key)->v.start) = U64_MAX - id; bch2_bset_init_first(b, &b->data->keys); bch2_btree_build_aux_trees(b); b->data->flags = 0; - b->data->min_key = POS_MIN; - b->data->max_key = POS_MAX; + btree_set_min(b, POS_MIN); + btree_set_max(b, POS_MAX); b->data->format = bch2_btree_calc_format(b); btree_node_set_format(b, b->data->format); - ret = bch2_btree_node_hash_insert(&c->btree_cache, b, b->level, b->btree_id); + ret = bch2_btree_node_hash_insert(&c->btree_cache, b, + b->c.level, b->c.btree_id); BUG_ON(ret); - __bch2_btree_set_root_inmem(c, b); + bch2_btree_set_root_inmem(c, b); - six_unlock_write(&b->lock); - six_unlock_intent(&b->lock); + six_unlock_write(&b->c.lock); + six_unlock_intent(&b->c.lock); } ssize_t bch2_btree_updates_print(struct bch_fs *c, char *buf) @@ -2243,3 +1999,75 @@ size_t bch2_btree_interior_updates_nr_pending(struct bch_fs *c) return ret; } + +void bch2_journal_entries_to_btree_roots(struct bch_fs *c, struct jset *jset) +{ + struct btree_root *r; + struct jset_entry *entry; + + mutex_lock(&c->btree_root_lock); + + vstruct_for_each(jset, entry) + if (entry->type == BCH_JSET_ENTRY_btree_root) { + r = &c->btree_roots[entry->btree_id]; + r->level = entry->level; + r->alive = true; + bkey_copy(&r->key, &entry->start[0]); + } + + mutex_unlock(&c->btree_root_lock); +} + +struct jset_entry * +bch2_btree_roots_to_journal_entries(struct bch_fs *c, + struct jset_entry *start, + struct jset_entry *end) +{ + struct jset_entry *entry; + unsigned long have = 0; + unsigned i; + + for (entry = start; entry < end; entry = vstruct_next(entry)) + if (entry->type == BCH_JSET_ENTRY_btree_root) + __set_bit(entry->btree_id, &have); + + mutex_lock(&c->btree_root_lock); + + for (i = 0; i < BTREE_ID_NR; i++) + if (c->btree_roots[i].alive && !test_bit(i, &have)) { + journal_entry_set(end, + BCH_JSET_ENTRY_btree_root, + i, c->btree_roots[i].level, + &c->btree_roots[i].key, + c->btree_roots[i].key.u64s); + end = vstruct_next(end); + } + + mutex_unlock(&c->btree_root_lock); + + return end; +} + +void bch2_fs_btree_interior_update_exit(struct bch_fs *c) +{ + if (c->btree_interior_update_worker) + destroy_workqueue(c->btree_interior_update_worker); + mempool_exit(&c->btree_interior_update_pool); +} + +int bch2_fs_btree_interior_update_init(struct bch_fs *c) +{ + mutex_init(&c->btree_reserve_cache_lock); + INIT_LIST_HEAD(&c->btree_interior_update_list); + INIT_LIST_HEAD(&c->btree_interior_updates_unwritten); + mutex_init(&c->btree_interior_update_lock); + INIT_WORK(&c->btree_interior_update_work, btree_interior_update_work); + + c->btree_interior_update_worker = + alloc_workqueue("btree_update", WQ_UNBOUND|WQ_MEM_RECLAIM, 1); + if (!c->btree_interior_update_worker) + return -ENOMEM; + + return mempool_init_kmalloc_pool(&c->btree_interior_update_pool, 1, + sizeof(struct btree_update)); +}