X-Git-Url: https://git.sesse.net/?a=blobdiff_plain;f=libbcachefs%2Fbtree_update_interior.c;h=f1e796336dde3379e8577d5d3318b3919d080643;hb=da6a35689518599b381c285cd9505ab8d58f7c73;hp=02f19146326287fb849c495a092d4acaf5e3098f;hpb=2286a79ccb086d8e850cb22df9d02d71d3aacc43;p=bcachefs-tools-debian diff --git a/libbcachefs/btree_update_interior.c b/libbcachefs/btree_update_interior.c index 02f1914..f1e7963 100644 --- a/libbcachefs/btree_update_interior.c +++ b/libbcachefs/btree_update_interior.c @@ -11,60 +11,88 @@ #include "btree_iter.h" #include "btree_locking.h" #include "buckets.h" +#include "error.h" #include "extents.h" #include "journal.h" #include "journal_reclaim.h" #include "keylist.h" +#include "recovery.h" #include "replicas.h" #include "super-io.h" #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); +static int bch2_btree_insert_node(struct btree_update *, struct btree_trans *, + struct btree_path *, struct btree *, + struct keylist *, unsigned); +static void bch2_btree_update_add_new_node(struct btree_update *, struct btree *); + +static struct btree_path *get_unlocked_mut_path(struct btree_trans *trans, + enum btree_id btree_id, + unsigned level, + struct bpos pos) +{ + struct btree_path *path; + + path = bch2_path_get(trans, btree_id, pos, level + 1, level, + BTREE_ITER_NOPRESERVE| + BTREE_ITER_INTENT, _RET_IP_); + path = bch2_btree_path_make_mut(trans, path, true, _RET_IP_); + bch2_btree_path_downgrade(trans, path); + __bch2_btree_path_unlock(trans, path); + return path; +} /* Debug code: */ -static void btree_node_interior_verify(struct btree *b) +/* + * Verify that child nodes correctly span parent node's range: + */ +static void btree_node_interior_verify(struct bch_fs *c, 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; + struct printbuf buf1 = PRINTBUF, buf2 = PRINTBUF; - 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)); + if (!test_bit(JOURNAL_REPLAY_DONE, &c->journal.flags)) + return; - BUG_ON((bch2_btree_node_iter_advance(&iter, b), - !bch2_btree_node_iter_end(&iter))); -#else - const char *msg; + bch2_btree_node_iter_init_from_start(&iter, b); - msg = "not found"; - k = bch2_btree_node_iter_peek(&iter, b); - if (!k) - goto err; + 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 = "isn't what it should be"; - if (bkey_cmp_left_packed(b, k, &b->key.k.p)) - goto err; + if (!bpos_eq(next_node, bp.v->min_key)) { + bch2_dump_btree_node(c, b); + bch2_bpos_to_text(&buf1, next_node); + bch2_bpos_to_text(&buf2, bp.v->min_key); + panic("expected next min_key %s got %s\n", buf1.buf, buf2.buf); + } - bch2_btree_node_iter_advance(&iter, b); + bch2_btree_node_iter_advance(&iter, b); - 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(); + if (bch2_btree_node_iter_end(&iter)) { + if (!bpos_eq(k.k->p, b->key.k.p)) { + bch2_dump_btree_node(c, b); + bch2_bpos_to_text(&buf1, b->key.k.p); + bch2_bpos_to_text(&buf2, k.k->p); + panic("expected end %s got %s\n", buf1.buf, buf2.buf); + } + break; + } + + next_node = bpos_successor(k.k->p); + } #endif } @@ -76,11 +104,9 @@ void __bch2_btree_calc_format(struct bkey_format_state *s, struct btree *b) struct bset_tree *t; struct bkey uk; - bch2_bkey_format_add_pos(s, b->data->min_key); - for_each_bset(b, t) bset_tree_for_each_key(b, t, k) - if (!bkey_whiteout(k)) { + if (!bkey_deleted(k)) { uk = bkey_unpack_key(b, k); bch2_bkey_format_add_key(s, &uk); } @@ -91,6 +117,8 @@ static struct bkey_format bch2_btree_calc_format(struct btree *b) struct bkey_format_state s; bch2_bkey_format_init(&s); + bch2_bkey_format_add_pos(&s, b->data->min_key); + bch2_bkey_format_add_pos(&s, b->data->max_key); __bch2_btree_calc_format(&s, b); return bch2_bkey_format_done(&s); @@ -129,78 +157,11 @@ 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); + trace_and_count(c, btree_node_free, c, b); + BUG_ON(btree_node_write_blocked(b)); BUG_ON(btree_node_dirty(b)); BUG_ON(btree_node_need_write(b)); BUG_ON(b == btree_node_root(c, b)); @@ -210,92 +171,90 @@ static void __btree_node_free(struct bch_fs *c, struct btree *b) clear_btree_node_noevict(b); - bch2_btree_node_hash_remove(&c->btree_cache, b); - mutex_lock(&c->btree_cache.lock); list_move(&b->list, &c->btree_cache.freeable); mutex_unlock(&c->btree_cache.lock); } -void bch2_btree_node_free_never_inserted(struct bch_fs *c, struct btree *b) +static void bch2_btree_node_free_inmem(struct btree_trans *trans, + struct btree_path *path, + struct btree *b) { - struct open_buckets ob = b->ob; - - btree_update_drop_new_node(c, b); - - b->ob.nr = 0; + struct bch_fs *c = trans->c; + unsigned level = b->c.level; - clear_btree_node_dirty(b); - - btree_node_lock_type(c, b, SIX_LOCK_write); + bch2_btree_node_lock_write_nofail(trans, path, &b->c); + bch2_btree_node_hash_remove(&c->btree_cache, b); __btree_node_free(c, b); - six_unlock_write(&b->lock); + six_unlock_write(&b->c.lock); + mark_btree_node_locked_noreset(path, level, SIX_LOCK_intent); - bch2_open_buckets_put(c, &ob); + trans_for_each_path(trans, path) + if (path->l[level].b == b) { + btree_node_unlock(trans, path, level); + path->l[level].b = ERR_PTR(-BCH_ERR_no_btree_node_init); + } } -void bch2_btree_node_free_inmem(struct bch_fs *c, struct btree *b, - struct btree_iter *iter) +static void bch2_btree_node_free_never_used(struct btree_update *as, + struct btree_trans *trans, + struct btree *b) { - struct btree_iter *linked; + struct bch_fs *c = as->c; + struct prealloc_nodes *p = &as->prealloc_nodes[b->c.lock.readers != NULL]; + struct btree_path *path; + unsigned level = b->c.level; - trans_for_each_iter(iter->trans, linked) - BUG_ON(linked->l[b->level].b == b); + BUG_ON(!list_empty(&b->write_blocked)); + BUG_ON(b->will_make_reachable != (1UL|(unsigned long) as)); - /* - * 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); + b->will_make_reachable = 0; + closure_put(&as->cl); - six_lock_write(&b->lock); - __btree_node_free(c, b); - six_unlock_write(&b->lock); - six_unlock_intent(&b->lock); -} + clear_btree_node_will_make_reachable(b); + clear_btree_node_accessed(b); + clear_btree_node_dirty_acct(c, b); + clear_btree_node_need_write(b); -static void bch2_btree_node_free_ondisk(struct bch_fs *c, - struct pending_btree_node_free *pending) -{ - BUG_ON(!pending->index_update_done); + mutex_lock(&c->btree_cache.lock); + list_del_init(&b->list); + bch2_btree_node_hash_remove(&c->btree_cache, b); + mutex_unlock(&c->btree_cache.lock); + + BUG_ON(p->nr >= ARRAY_SIZE(p->b)); + p->b[p->nr++] = b; - bch2_mark_key(c, bkey_i_to_s_c(&pending->key), - 0, 0, NULL, 0, BTREE_TRIGGER_OVERWRITE); + six_unlock_intent(&b->c.lock); - 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); + trans_for_each_path(trans, path) + if (path->l[level].b == b) { + btree_node_unlock(trans, path, level); + path->l[level].b = ERR_PTR(-BCH_ERR_no_btree_node_init); + } } -static struct btree *__bch2_btree_node_alloc(struct bch_fs *c, +static struct btree *__bch2_btree_node_alloc(struct btree_trans *trans, struct disk_reservation *res, struct closure *cl, + bool interior_node, unsigned flags) { + struct bch_fs *c = trans->c; struct write_point *wp; struct btree *b; - BKEY_PADDED(k) tmp; + __BKEY_PADDED(k, BKEY_BTREE_PTR_VAL_U64s_MAX) tmp; struct open_buckets ob = { .nr = 0 }; struct bch_devs_list devs_have = (struct bch_devs_list) { 0 }; unsigned nr_reserve; enum alloc_reserve alloc_reserve; + int ret; - if (flags & BTREE_INSERT_USE_ALLOC_RESERVE) { + if (flags & BTREE_INSERT_USE_RESERVE) { nr_reserve = 0; - alloc_reserve = RESERVE_ALLOC; - } else if (flags & BTREE_INSERT_USE_RESERVE) { - nr_reserve = BTREE_NODE_RESERVE / 2; - alloc_reserve = RESERVE_BTREE; + alloc_reserve = RESERVE_btree_movinggc; } else { nr_reserve = BTREE_NODE_RESERVE; - alloc_reserve = RESERVE_NONE; + alloc_reserve = RESERVE_btree; } mutex_lock(&c->btree_reserve_cache_lock); @@ -311,38 +270,39 @@ static struct btree *__bch2_btree_node_alloc(struct bch_fs *c, mutex_unlock(&c->btree_reserve_cache_lock); retry: - wp = bch2_alloc_sectors_start(c, c->opts.foreground_target, 0, + ret = bch2_alloc_sectors_start_trans(trans, + c->opts.metadata_target ?: + c->opts.foreground_target, + 0, writepoint_ptr(&c->btree_write_point), &devs_have, res->nr_replicas, c->opts.metadata_replicas_required, - alloc_reserve, 0, cl); - if (IS_ERR(wp)) - return ERR_CAST(wp); + alloc_reserve, 0, cl, &wp); + if (unlikely(ret)) + return ERR_PTR(ret); - if (wp->sectors_free < c->opts.btree_node_size) { + if (wp->sectors_free < btree_sectors(c)) { struct open_bucket *ob; unsigned i; open_bucket_for_each(c, &wp->ptrs, ob, i) - if (ob->sectors_free < c->opts.btree_node_size) + if (ob->sectors_free < btree_sectors(c)) ob->sectors_free = 0; bch2_alloc_sectors_done(c, wp); goto retry; } - 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); + bkey_btree_ptr_v2_init(&tmp.k); + bch2_alloc_sectors_append_ptrs(c, wp, &tmp.k, btree_sectors(c), false); bch2_open_bucket_get(c, wp, &ob); bch2_alloc_sectors_done(c, wp); mem_alloc: - b = bch2_btree_node_mem_alloc(c); + b = bch2_btree_node_mem_alloc(c, interior_node); + six_unlock_write(&b->c.lock); + six_unlock_intent(&b->c.lock); /* we hold cannibalize_lock: */ BUG_ON(IS_ERR(b)); @@ -354,31 +314,38 @@ mem_alloc: return b; } -static struct btree *bch2_btree_node_alloc(struct btree_update *as, unsigned level) +static struct btree *bch2_btree_node_alloc(struct btree_update *as, + struct btree_trans *trans, + unsigned level) { struct bch_fs *c = as->c; struct btree *b; + struct prealloc_nodes *p = &as->prealloc_nodes[!!level]; int ret; BUG_ON(level >= BTREE_MAX_DEPTH); - BUG_ON(!as->reserve->nr); + BUG_ON(!p->nr); - b = as->reserve->b[--as->reserve->nr]; + b = p->b[--p->nr]; + + btree_node_lock_nopath_nofail(trans, &b->c, SIX_LOCK_intent); + btree_node_lock_nopath_nofail(trans, &b->c, SIX_LOCK_write); set_btree_node_accessed(b); - set_btree_node_dirty(b); + set_btree_node_dirty_acct(c, b); set_btree_node_need_write(b); bch2_bset_init_first(b, &b->data->keys); - b->level = level; - b->btree_id = as->btree_id; + b->c.level = level; + b->c.btree_id = as->btree_id; + b->version_ondisk = c->sb.version; memset(&b->nr, 0, sizeof(b->nr)); b->data->magic = cpu_to_le64(bset_magic(c)); + memset(&b->data->_ptr, 0, sizeof(b->data->_ptr)); b->data->flags = 0; SET_BTREE_NODE_ID(b->data, as->btree_id); SET_BTREE_NODE_LEVEL(b->data, level); - 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); @@ -386,24 +353,16 @@ static struct btree *bch2_btree_node_alloc(struct btree_update *as, unsigned lev 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); - - if (btree_node_is_extents(b) && - !BTREE_NODE_NEW_EXTENT_OVERWRITE(b->data)) - set_btree_node_old_extent_overwrite(b); + SET_BTREE_NODE_NEW_EXTENT_OVERWRITE(b->data, true); 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); + trace_and_count(c, btree_node_alloc, c, b); return b; } @@ -420,13 +379,19 @@ static void btree_set_max(struct btree *b, struct bpos pos) b->data->max_key = pos; } -struct btree *__bch2_btree_node_alloc_replacement(struct btree_update *as, - struct btree *b, - struct bkey_format format) +static struct btree *bch2_btree_node_alloc_replacement(struct btree_update *as, + struct btree_trans *trans, + struct btree *b) { - struct btree *n; + struct btree *n = bch2_btree_node_alloc(as, trans, b->c.level); + struct bkey_format format = bch2_btree_calc_format(b); - n = bch2_btree_node_alloc(as, b->level); + /* + * The keys might expand with the new format - if they wouldn't fit in + * the btree node anymore, use the old format for now: + */ + if (!bch2_btree_node_format_fits(as->c, b, &format)) + format = b->format; SET_BTREE_NODE_SEQ(n->data, BTREE_NODE_SEQ(b->data) + 1); @@ -439,355 +404,394 @@ struct btree *__bch2_btree_node_alloc_replacement(struct btree_update *as, bch2_btree_sort_into(as->c, n, b); btree_node_reset_sib_u64s(n); - - n->key.k.p = b->key.k.p; return n; } -static struct btree *bch2_btree_node_alloc_replacement(struct btree_update *as, - struct btree *b) -{ - struct bkey_format new_f = bch2_btree_calc_format(b); - - /* - * The keys might expand with the new format - if they wouldn't fit in - * the btree node anymore, use the old format for now: - */ - if (!bch2_btree_node_format_fits(as->c, b, &new_f)) - new_f = b->format; - - return __bch2_btree_node_alloc_replacement(as, b, new_f); -} - -static struct btree *__btree_root_alloc(struct btree_update *as, unsigned level) +static struct btree *__btree_root_alloc(struct btree_update *as, + struct btree_trans *trans, unsigned level) { - struct btree *b = bch2_btree_node_alloc(as, level); + struct btree *b = bch2_btree_node_alloc(as, trans, level); btree_set_min(b, POS_MIN); - btree_set_max(b, POS_MAX); + btree_set_max(b, SPOS_MAX); b->data->format = bch2_btree_calc_format(b); btree_node_set_format(b, b->data->format); bch2_btree_build_aux_trees(b); - six_unlock_write(&b->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, struct btree_trans *trans) { - bch2_disk_reservation_put(c, &reserve->disk_res); - - mutex_lock(&c->btree_reserve_cache_lock); - - while (reserve->nr) { - struct btree *b = reserve->b[--reserve->nr]; - - six_unlock_write(&b->lock); + struct bch_fs *c = as->c; + struct prealloc_nodes *p; + + for (p = as->prealloc_nodes; + p < as->prealloc_nodes + ARRAY_SIZE(as->prealloc_nodes); + p++) { + while (p->nr) { + struct btree *b = p->b[--p->nr]; + + mutex_lock(&c->btree_reserve_cache_lock); + + if (c->btree_reserve_cache_nr < + ARRAY_SIZE(c->btree_reserve_cache)) { + struct btree_alloc *a = + &c->btree_reserve_cache[c->btree_reserve_cache_nr++]; + + a->ob = b->ob; + b->ob.nr = 0; + bkey_copy(&a->k, &b->key); + } else { + bch2_open_buckets_put(c, &b->ob); + } - if (c->btree_reserve_cache_nr < - ARRAY_SIZE(c->btree_reserve_cache)) { - struct btree_alloc *a = - &c->btree_reserve_cache[c->btree_reserve_cache_nr++]; + mutex_unlock(&c->btree_reserve_cache_lock); - a->ob = b->ob; - b->ob.nr = 0; - bkey_copy(&a->k, &b->key); - } else { - bch2_open_buckets_put(c, &b->ob); + btree_node_lock_nopath_nofail(trans, &b->c, SIX_LOCK_intent); + btree_node_lock_nopath_nofail(trans, &b->c, SIX_LOCK_write); + __btree_node_free(c, b); + six_unlock_write(&b->c.lock); + six_unlock_intent(&b->c.lock); } - - btree_node_lock_type(c, b, SIX_LOCK_write); - __btree_node_free(c, b); - six_unlock_write(&b->lock); - - six_unlock_intent(&b->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_trans *trans, + struct btree_update *as, + unsigned nr_nodes[2], + 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); + unsigned interior; + int ret = 0; - BUG_ON(nr_nodes > BTREE_RESERVE_MAX); + BUG_ON(nr_nodes[0] + nr_nodes[1] > BTREE_RESERVE_MAX); /* * Protects reaping from the btree node cache and using the btree node * open bucket reserve: + * + * BTREE_INSERT_NOWAIT only applies to btree node allocation, not + * blocking on this lock: */ ret = bch2_btree_cache_cannibalize_lock(c, cl); - if (ret) { - bch2_disk_reservation_put(c, &disk_res); - return ERR_PTR(ret); - } + if (ret) + return ret; - reserve = mempool_alloc(&c->btree_reserve_pool, GFP_NOIO); + for (interior = 0; interior < 2; interior++) { + struct prealloc_nodes *p = as->prealloc_nodes + interior; - reserve->disk_res = disk_res; - reserve->nr = 0; + while (p->nr < nr_nodes[interior]) { + b = __bch2_btree_node_alloc(trans, &as->disk_res, + flags & BTREE_INSERT_NOWAIT ? NULL : cl, + interior, flags); + if (IS_ERR(b)) { + ret = PTR_ERR(b); + goto err; + } - while (reserve->nr < nr_nodes) { - b = __bch2_btree_node_alloc(c, &disk_res, - flags & BTREE_INSERT_NOWAIT - ? NULL : cl, flags); - if (IS_ERR(b)) { - ret = PTR_ERR(b); - goto err_free; + p->b[p->nr++] = b; } - - ret = bch2_mark_bkey_replicas(c, bkey_i_to_s_c(&b->key)); - if (ret) - goto err_free; - - reserve->b[reserve->nr++] = b; } - - bch2_btree_cache_cannibalize_unlock(c); - return reserve; -err_free: - bch2_btree_reserve_put(c, reserve); +err: 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 */ -static void bch2_btree_update_free(struct btree_update *as) +static void bch2_btree_update_free(struct btree_update *as, struct btree_trans *trans) { struct bch_fs *c = as->c; - bch2_journal_pin_flush(&c->journal, &as->journal); + if (as->took_gc_lock) + up_read(&c->gc_lock); + as->took_gc_lock = false; - BUG_ON(as->nr_new_nodes); - BUG_ON(as->nr_pending); + bch2_journal_preres_put(&c->journal, &as->journal_preres); + + 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, trans); - if (as->reserve) - bch2_btree_reserve_put(c, as->reserve); + bch2_time_stats_update(&c->times[BCH_TIME_btree_interior_update_total], + as->start_time); mutex_lock(&c->btree_interior_update_lock); + list_del(&as->unwritten_list); list_del(&as->list); closure_debug_destroy(&as->cl); mempool_free(as, &c->btree_interior_update_pool); + /* + * Have to do the wakeup with btree_interior_update_lock still held, + * since being on btree_interior_update_list is our ref on @c: + */ 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_add_key(struct btree_update *as, + struct keylist *keys, struct btree *b) { - struct btree_update *as = container_of(cl, struct btree_update, cl); - struct bch_fs *c = as->c; + struct bkey_i *k = &b->key; - bch2_journal_pin_drop(&c->journal, &as->journal); + BUG_ON(bch2_keylist_u64s(keys) + k->k.u64s > + ARRAY_SIZE(as->_old_keys)); - mutex_lock(&c->btree_interior_update_lock); + bkey_copy(keys->top, k); + bkey_i_to_btree_ptr_v2(keys->top)->v.mem_ptr = b->c.level + 1; - while (as->nr_new_nodes) { - struct btree *b = as->new_nodes[--as->nr_new_nodes]; + bch2_keylist_push(keys); +} - BUG_ON(b->will_make_reachable != (unsigned long) as); - b->will_make_reachable = 0; - mutex_unlock(&c->btree_interior_update_lock); +/* + * 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 bkey_i *k; + int ret; - /* - * 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); - } + ret = darray_make_room(&trans->extra_journal_entries, as->journal_u64s); + if (ret) + return ret; - while (as->nr_pending) - bch2_btree_node_free_ondisk(c, &as->pending[--as->nr_pending]); + memcpy(&darray_top(trans->extra_journal_entries), + as->journal_entries, + as->journal_u64s * sizeof(u64)); + trans->extra_journal_entries.nr += as->journal_u64s; - mutex_unlock(&c->btree_interior_update_lock); + trans->journal_pin = &as->journal; - closure_wake_up(&as->wait); + for_each_keylist_key(&as->old_keys, k) { + unsigned level = bkey_i_to_btree_ptr_v2(k)->v.mem_ptr; - bch2_btree_update_free(as); -} + ret = bch2_trans_mark_old(trans, as->btree_id, level, bkey_i_to_s_c(k), 0); + if (ret) + return ret; + } -static void btree_update_wait_on_journal(struct closure *cl) -{ - struct btree_update *as = container_of(cl, struct btree_update, cl); - struct bch_fs *c = as->c; - int ret; + for_each_keylist_key(&as->new_keys, k) { + unsigned level = bkey_i_to_btree_ptr_v2(k)->v.mem_ptr; - 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; + ret = bch2_trans_mark_new(trans, as->btree_id, level, k, 0); + 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); + 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_trans trans; + u64 journal_seq = 0; + unsigned i; + int ret; + bch2_trans_init(&trans, c, 0, 512); /* - * 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. + * If we're already in an error state, it might be because a btree node + * was never written, and we might be trying to free that same btree + * node here, but it won't have been marked as allocated and we'll see + * spurious disk usage inconsistencies in the transactional part below + * if we don't skip it: */ - mutex_lock(&c->btree_interior_update_lock); - as->nodes_written = true; -retry: - as = list_first_entry_or_null(&c->btree_interior_updates_unwritten, - struct btree_update, unwritten_list); - if (!as || !as->nodes_written) { - mutex_unlock(&c->btree_interior_update_lock); - return; - } + ret = bch2_journal_error(&c->journal); + if (ret) + goto err; - 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); - mutex_lock(&c->btree_interior_update_lock); - goto retry; - } + /* + * Wait for any in flight writes to finish before we free the old nodes + * on disk: + */ + for (i = 0; i < as->nr_old_nodes; i++) { + __le64 seq; - BUG_ON(!btree_node_dirty(b)); - closure_wait(&btree_current_write(b)->wait, &as->cl); + b = as->old_nodes[i]; - list_del(&as->write_blocked_list); + btree_node_lock_nopath_nofail(&trans, &b->c, SIX_LOCK_read); + seq = b->data ? b->data->keys.seq : 0; + six_unlock_read(&b->c.lock); - /* - * for flush_held_btree_writes() waiting on updates to flush or - * nodes to be writeable: - */ - closure_wake_up(&c->btree_interior_update_wait); + if (seq == as->old_nodes_seq[i]) + wait_on_bit_io(&b->flags, BTREE_NODE_write_in_flight_inner, + TASK_UNINTERRUPTIBLE); + } - list_del(&as->unwritten_list); - mutex_unlock(&c->btree_interior_update_lock); + /* + * 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. + */ + + /* + * 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 = commit_do(&trans, &as->disk_res, &journal_seq, + BTREE_INSERT_NOFAIL| + BTREE_INSERT_NOCHECK_RW| + BTREE_INSERT_USE_RESERVE| + BTREE_INSERT_JOURNAL_RECLAIM| + JOURNAL_WATERMARK_reserved, + btree_update_nodes_written_trans(&trans, as)); + bch2_trans_unlock(&trans); + + bch2_fs_fatal_err_on(ret && !bch2_journal_error(&c->journal), c, + "%s(): error %s", __func__, bch2_err_str(ret)); +err: + if (as->b) { + struct btree_path *path; + b = as->b; + path = get_unlocked_mut_path(&trans, as->btree_id, b->c.level, b->key.k.p); /* - * b->write_blocked prevented it from being written, so - * write it now if it needs to be written: + * @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: */ - bch2_btree_node_write_cond(c, b, true); - six_unlock_read(&b->lock); - continue_at(&as->cl, btree_update_nodes_reachable, system_wq); - break; - 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: + * Ensure transaction is unlocked before using + * btree_node_lock_nopath() (the use of which is always suspect, + * we need to work on removing this in the future) + * + * It should be, but get_unlocked_mut_path() -> bch2_path_get() + * calls bch2_path_upgrade(), before we call path_make_mut(), so + * we may rarely end up with a locked path besides the one we + * have here: */ - closure_put(&as->parent_as->cl); + bch2_trans_unlock(&trans); + btree_node_lock_nopath_nofail(&trans, &b->c, SIX_LOCK_intent); + mark_btree_node_locked(&trans, path, b->c.level, SIX_LOCK_intent); + bch2_btree_path_level_init(&trans, path, b); + + bch2_btree_node_lock_write_nofail(&trans, path, &b->c); + + mutex_lock(&c->btree_interior_update_lock); + + list_del(&as->write_blocked_list); + if (list_empty(&b->write_blocked)) + clear_btree_node_write_blocked(b); /* - * and then we have to wait on that btree_update to finish: + * Node might have been freed, recheck under + * btree_interior_update_lock: */ - closure_wait(&as->parent_as->wait, &as->cl); + if (as->b == b) { + struct bset *i = btree_bset_last(b); + + BUG_ON(!b->c.level); + BUG_ON(!btree_node_dirty(b)); + + if (!ret) { + i->journal_seq = cpu_to_le64( + max(journal_seq, + le64_to_cpu(i->journal_seq))); + + bch2_btree_add_journal_pin(c, b, journal_seq); + } else { + /* + * If we didn't get a journal sequence number we + * can't write this btree node, because recovery + * won't know to ignore this write: + */ + set_btree_node_never_write(b); + } + } - list_del(&as->unwritten_list); mutex_unlock(&c->btree_interior_update_lock); - continue_at(&as->cl, btree_update_nodes_reachable, system_wq); - break; + mark_btree_node_locked_noreset(path, b->c.level, SIX_LOCK_intent); + 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); + btree_node_unlock(&trans, path, b->c.level); + bch2_path_put(&trans, path, true); + } - 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); - mutex_lock(&c->btree_interior_update_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); + BUG_ON(b->will_make_reachable != (unsigned long) as); + b->will_make_reachable = 0; + clear_btree_node_will_make_reachable(b); + } + mutex_unlock(&c->btree_interior_update_lock); - list_del(&as->unwritten_list); - mutex_unlock(&c->btree_interior_update_lock); + for (i = 0; i < as->nr_new_nodes; i++) { + b = as->new_nodes[i]; - /* - * 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); + btree_node_lock_nopath_nofail(&trans, &b->c, SIX_LOCK_read); + btree_node_write_if_need(c, b, SIX_LOCK_read); + six_unlock_read(&b->c.lock); + } + + 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, &trans); + bch2_trans_exit(&trans); +} + +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); - as->journal_seq = bch2_journal_last_unwritten_seq(&c->journal); + if (!as) + break; - btree_update_wait_on_journal(&as->cl); - 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); - goto retry; + as->nodes_written = true; + mutex_unlock(&c->btree_interior_update_lock); + + queue_work(c->btree_interior_update_worker, &c->btree_interior_update_work); } /* @@ -803,128 +807,95 @@ static void btree_update_updated_node(struct btree_update *as, struct btree *b) BUG_ON(as->mode != BTREE_INTERIOR_NO_UPDATE); BUG_ON(!btree_node_dirty(b)); + BUG_ON(!b->c.level); + + as->mode = BTREE_INTERIOR_UPDATING_NODE; + as->b = b; - as->mode = BTREE_INTERIOR_UPDATING_NODE; - as->b = b; + set_btree_node_write_blocked(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, struct btree_update *child) { 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 - * _before_ the new nodes are technically reachable; see - * btree_update_nodes_written(). - * - * This goes for journal pins that are recursively blocked on us - so, - * just transfer the journal pin to the new interior update so - * btree_update_nodes_written() can drop it. - */ - bch2_journal_pin_copy(&c->journal, &as->journal, - &child->journal, interior_update_flush); - bch2_journal_pin_drop(&c->journal, &child->journal); - as->journal_seq = max(as->journal_seq, child->journal_seq); + bch2_journal_pin_copy(&c->journal, &as->journal, &child->journal, NULL); } -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); - list_add_tail(&as->unwritten_list, &c->btree_interior_updates_unwritten); 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: + */ +static 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); as->new_nodes[as->nr_new_nodes++] = b; b->will_make_reachable = 1UL|(unsigned long) as; + set_btree_node_will_make_reachable(b); - closure_get(&as->cl); mutex_unlock(&c->btree_interior_update_lock); + + btree_update_add_key(as, &as->new_keys, b); + + if (b->key.k.type == KEY_TYPE_btree_ptr_v2) { + unsigned bytes = vstruct_end(&b->data->keys) - (void *) b->data; + unsigned sectors = round_up(bytes, block_bytes(c)) >> 9; + + bkey_i_to_btree_ptr_v2(&b->key)->v.sectors_written = + cpu_to_le16(sectors); + } } +/* + * 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; @@ -932,7 +903,13 @@ 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); + clear_btree_node_will_make_reachable(b); as = (struct btree_update *) (v & ~1UL); if (!as) { @@ -953,25 +930,11 @@ found: closure_put(&as->cl); } -static void btree_interior_update_add_node_reference(struct btree_update *as, - struct btree *b) +static 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]; } /* @@ -979,34 +942,18 @@ static void btree_interior_update_add_node_reference(struct btree_update *as, * nodes and thus outstanding btree_updates - redirect @b's * btree_updates to point to this btree_update: */ -void bch2_btree_interior_update_will_free_node(struct btree_update *as, - struct btree *b) +static 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); /* @@ -1018,7 +965,7 @@ 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); /* @@ -1028,18 +975,9 @@ void bch2_btree_interior_update_will_free_node(struct btree_update *as, closure_wake_up(&c->btree_interior_update_wait); } - clear_btree_node_dirty(b); + clear_btree_node_dirty_acct(c, 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); + clear_btree_node_write_blocked(b); /* * Does this node have unwritten data that has a pin on the journal? @@ -1049,61 +987,207 @@ 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_copy(&c->journal, &as->journal, - &w->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_copy(&c->journal, &as->journal, - &w->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_add_key(as, &as->old_keys, b); + + as->old_nodes[as->nr_old_nodes] = b; + as->old_nodes_seq[as->nr_old_nodes] = b->data->keys.seq; + as->nr_old_nodes++; } -void bch2_btree_update_done(struct btree_update *as) +static void bch2_btree_update_done(struct btree_update *as, struct btree_trans *trans) { + struct bch_fs *c = as->c; + u64 start_time = as->start_time; + BUG_ON(as->mode == BTREE_INTERIOR_NO_UPDATE); - bch2_btree_reserve_put(as->c, as->reserve); - as->reserve = NULL; + if (as->took_gc_lock) + up_read(&as->c->gc_lock); + as->took_gc_lock = false; + + bch2_btree_reserve_put(as, trans); - continue_at(&as->cl, btree_update_nodes_written, system_freezable_wq); + continue_at(&as->cl, btree_update_set_nodes_written, + as->c->btree_interior_update_worker); + + bch2_time_stats_update(&c->times[BCH_TIME_btree_interior_update_foreground], + start_time); } -struct btree_update * -bch2_btree_update_start(struct bch_fs *c, enum btree_id id, - unsigned nr_nodes, unsigned flags, - struct closure *cl) +static struct btree_update * +bch2_btree_update_start(struct btree_trans *trans, struct btree_path *path, + unsigned level, bool split, unsigned flags) { - struct btree_reserve *reserve; + struct bch_fs *c = trans->c; struct btree_update *as; + u64 start_time = local_clock(); + int disk_res_flags = (flags & BTREE_INSERT_NOFAIL) + ? BCH_DISK_RESERVATION_NOFAIL : 0; + unsigned nr_nodes[2] = { 0, 0 }; + unsigned update_level = level; + int journal_flags = flags & JOURNAL_WATERMARK_MASK; + int ret = 0; + u32 restart_count = trans->restart_count; + + BUG_ON(!path->should_be_locked); + + if (flags & BTREE_INSERT_JOURNAL_RECLAIM) + journal_flags |= JOURNAL_RES_GET_NONBLOCK; + + while (1) { + nr_nodes[!!update_level] += 1 + split; + update_level++; + + ret = bch2_btree_path_upgrade(trans, path, update_level + 1); + if (ret) + return ERR_PTR(ret); - reserve = bch2_btree_reserve_get(c, nr_nodes, flags, cl); - if (IS_ERR(reserve)) - return ERR_CAST(reserve); + if (!btree_path_node(path, update_level)) { + /* Allocating new root? */ + nr_nodes[1] += split; + update_level = BTREE_MAX_DEPTH; + break; + } + + if (bch2_btree_node_insert_fits(c, path->l[update_level].b, + BKEY_BTREE_PTR_U64s_MAX * (1 + split))) + break; + + split = true; + } + + if (flags & BTREE_INSERT_GC_LOCK_HELD) + lockdep_assert_held(&c->gc_lock); + else if (!down_read_trylock(&c->gc_lock)) { + bch2_trans_unlock(trans); + down_read(&c->gc_lock); + ret = bch2_trans_relock(trans); + if (ret) { + up_read(&c->gc_lock); + return ERR_PTR(ret); + } + } as = mempool_alloc(&c->btree_interior_update_pool, GFP_NOIO); memset(as, 0, sizeof(*as)); closure_init(&as->cl, NULL); as->c = c; + as->start_time = start_time; as->mode = BTREE_INTERIOR_NO_UPDATE; - as->btree_id = id; - as->reserve = reserve; + as->took_gc_lock = !(flags & BTREE_INSERT_GC_LOCK_HELD); + as->btree_id = path->btree_id; + as->update_level = update_level; + 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); mutex_lock(&c->btree_interior_update_lock); list_add_tail(&as->list, &c->btree_interior_update_list); mutex_unlock(&c->btree_interior_update_lock); + /* + * We don't want to allocate if we're in an error state, that can cause + * deadlock on emergency shutdown due to open buckets getting stuck in + * the btree_reserve_cache after allocator shutdown has cleared it out. + * This check needs to come after adding us to the btree_interior_update + * list but before calling bch2_btree_reserve_get, to synchronize with + * __bch2_fs_read_only(). + */ + ret = bch2_journal_error(&c->journal); + if (ret) + goto err; + + ret = bch2_journal_preres_get(&c->journal, &as->journal_preres, + BTREE_UPDATE_JOURNAL_RES, + journal_flags|JOURNAL_RES_GET_NONBLOCK); + if (ret) { + bch2_trans_unlock(trans); + + if (flags & BTREE_INSERT_JOURNAL_RECLAIM) { + ret = -BCH_ERR_journal_reclaim_would_deadlock; + goto err; + } + + ret = bch2_journal_preres_get(&c->journal, &as->journal_preres, + BTREE_UPDATE_JOURNAL_RES, + journal_flags); + if (ret) { + trace_and_count(c, trans_restart_journal_preres_get, trans, _RET_IP_, journal_flags); + ret = btree_trans_restart(trans, BCH_ERR_transaction_restart_journal_preres_get); + goto err; + } + + ret = bch2_trans_relock(trans); + if (ret) + goto err; + } + + ret = bch2_disk_reservation_get(c, &as->disk_res, + (nr_nodes[0] + nr_nodes[1]) * btree_sectors(c), + c->opts.metadata_replicas, + disk_res_flags); + if (ret) + goto err; + + ret = bch2_btree_reserve_get(trans, as, nr_nodes, flags, NULL); + if (bch2_err_matches(ret, ENOSPC) || + bch2_err_matches(ret, ENOMEM)) { + struct closure cl; + + closure_init_stack(&cl); + + do { + ret = bch2_btree_reserve_get(trans, as, nr_nodes, flags, &cl); + + bch2_trans_unlock(trans); + closure_sync(&cl); + } while (bch2_err_matches(ret, BCH_ERR_operation_blocked)); + } + + if (ret) { + trace_and_count(c, btree_reserve_get_fail, trans->fn, + _RET_IP_, nr_nodes[0] + nr_nodes[1], ret); + goto err; + } + + ret = bch2_trans_relock(trans); + if (ret) + goto err; + + bch2_trans_verify_not_restarted(trans, restart_count); return as; +err: + bch2_btree_update_free(as, trans); + 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); @@ -1112,7 +1196,7 @@ 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; @@ -1121,54 +1205,6 @@ 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 * @@ -1181,15 +1217,15 @@ static void bch2_btree_set_root_ondisk(struct bch_fs *c, struct btree *b, int rw * is nothing new to be done. This just guarantees that there is a * journal write. */ -static void bch2_btree_set_root(struct btree_update *as, struct btree *b, - struct btree_iter *iter) +static void bch2_btree_set_root(struct btree_update *as, + struct btree_trans *trans, + struct btree_path *path, + struct btree *b) { struct bch_fs *c = as->c; struct btree *old; - trace_btree_set_root(c, b); - BUG_ON(!b->written && - !test_bit(BCH_FS_HOLD_BTREE_WRITES, &c->flags)); + trace_and_count(c, btree_node_set_root, c, b); old = btree_node_root(c, b); @@ -1197,11 +1233,11 @@ static void bch2_btree_set_root(struct btree_update *as, struct btree *b, * Ensure no one is using the old root while we switch to the * new root: */ - bch2_btree_node_lock_write(old, iter); + bch2_btree_node_lock_write_nofail(trans, path, &old->c); - 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: @@ -1210,148 +1246,189 @@ static void bch2_btree_set_root(struct btree_update *as, struct btree *b, * an intent lock on the new root, and any updates that would * depend on the new root would have to update the new root. */ - bch2_btree_node_unlock_write(old, iter); + bch2_btree_node_unlock_write(trans, path, old); } /* Interior node updates: */ -static void bch2_insert_fixup_btree_ptr(struct btree_update *as, struct btree *b, - struct btree_iter *iter, - struct bkey_i *insert, - struct btree_node_iter *node_iter) +static void bch2_insert_fixup_btree_ptr(struct btree_update *as, + struct btree_trans *trans, + struct btree_path *path, + struct btree *b, + struct btree_node_iter *node_iter, + struct bkey_i *insert) { 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); + struct printbuf buf = PRINTBUF; + unsigned long old, new, v; + + BUG_ON(insert->k.type == KEY_TYPE_btree_ptr_v2 && + !btree_ptr_sectors_written(insert)); + + if (unlikely(!test_bit(JOURNAL_REPLAY_DONE, &c->journal.flags))) + bch2_journal_key_overwritten(c, b->c.btree_id, b->c.level, insert->k.p); + + if (bch2_bkey_invalid(c, bkey_i_to_s_c(insert), + btree_node_type(b), WRITE, &buf) ?: + bch2_bkey_in_btree_node(b, bkey_i_to_s_c(insert), &buf)) { + printbuf_reset(&buf); + prt_printf(&buf, "inserting invalid bkey\n "); + bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(insert)); + prt_printf(&buf, "\n "); + bch2_bkey_invalid(c, bkey_i_to_s_c(insert), + btree_node_type(b), WRITE, &buf); + bch2_bkey_in_btree_node(b, bkey_i_to_s_c(insert), &buf); + + bch2_fs_inconsistent(c, "%s", buf.buf); + dump_stack(); + } - 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, 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_btree_bset_insert_key(trans, path, b, node_iter, insert); + set_btree_node_dirty_acct(c, b); - bch2_fs_usage_apply(c, fs_usage, &as->reserve->disk_res, 0); + v = READ_ONCE(b->flags); + do { + old = new = v; - bch2_fs_usage_scratch_put(c, fs_usage); - percpu_up_read(&c->mark_lock); - mutex_unlock(&c->btree_interior_update_lock); + new &= ~BTREE_WRITE_TYPE_MASK; + new |= BTREE_WRITE_interior; + new |= 1 << BTREE_NODE_need_write; + } while ((v = cmpxchg(&b->flags, old, new)) != old); - bch2_btree_bset_insert_key(iter, b, node_iter, insert); - set_btree_node_dirty(b); - set_btree_node_need_write(b); + printbuf_exit(&buf); +} + +static void +__bch2_btree_insert_keys_interior(struct btree_update *as, + struct btree_trans *trans, + struct btree_path *path, + struct btree *b, + struct btree_node_iter node_iter, + struct keylist *keys) +{ + struct bkey_i *insert = bch2_keylist_front(keys); + struct bkey_packed *k; + + BUG_ON(btree_node_type(b) != BKEY_TYPE_btree); + + while ((k = bch2_btree_node_iter_prev_all(&node_iter, b)) && + (bkey_cmp_left_packed(b, k, &insert->k.p) >= 0)) + ; + + while (!bch2_keylist_empty(keys)) { + struct bkey_i *k = bch2_keylist_front(keys); + + if (bpos_gt(k->k.p, b->key.k.p)) + break; + + bch2_insert_fixup_btree_ptr(as, trans, path, b, &node_iter, k); + bch2_keylist_pop_front(keys); + } } /* * Move keys from n1 (original replacement node, now lower node) to n2 (higher * node) */ -static struct btree *__btree_split_node(struct btree_update *as, - struct btree *n1, - struct btree_iter *iter) +static void __btree_split_node(struct btree_update *as, + struct btree_trans *trans, + struct btree *b, + struct btree *n[2]) { - size_t nr_packed = 0, nr_unpacked = 0; - struct btree *n2; - struct bset *set1, *set2; - struct bkey_packed *k, *prev = NULL; - - n2 = bch2_btree_node_alloc(as, n1->level); + struct bkey_packed *k; + struct bpos n1_pos = POS_MIN; + struct btree_node_iter iter; + struct bset *bsets[2]; + struct bkey_format_state format[2]; + struct bkey_packed *out[2]; + struct bkey uk; + unsigned u64s, n1_u64s = (b->nr.live_u64s * 3) / 5; + int i; - n2->data->max_key = n1->data->max_key; - n2->data->format = n1->format; - SET_BTREE_NODE_SEQ(n2->data, BTREE_NODE_SEQ(n1->data)); - n2->key.k.p = n1->key.k.p; + for (i = 0; i < 2; i++) { + BUG_ON(n[i]->nsets != 1); - btree_node_set_format(n2, n2->data->format); + bsets[i] = btree_bset_first(n[i]); + out[i] = bsets[i]->start; - set1 = btree_bset_first(n1); - set2 = btree_bset_first(n2); + SET_BTREE_NODE_SEQ(n[i]->data, BTREE_NODE_SEQ(b->data) + 1); + bch2_bkey_format_init(&format[i]); + } - /* - * Has to be a linear search because we don't have an auxiliary - * search tree yet - */ - k = set1->start; - while (1) { - struct bkey_packed *n = bkey_next_skip_noops(k, vstruct_last(set1)); + u64s = 0; + for_each_btree_node_key(b, k, &iter) { + if (bkey_deleted(k)) + continue; + + i = u64s >= n1_u64s; + u64s += k->u64s; + uk = bkey_unpack_key(b, k); + if (!i) + n1_pos = uk.p; + bch2_bkey_format_add_key(&format[i], &uk); + } - if (n == vstruct_last(set1)) - break; - if (k->_data - set1->_data >= (le16_to_cpu(set1->u64s) * 3) / 5) - break; + btree_set_min(n[0], b->data->min_key); + btree_set_max(n[0], n1_pos); + btree_set_min(n[1], bpos_successor(n1_pos)); + btree_set_max(n[1], b->data->max_key); - if (bkey_packed(k)) - nr_packed++; - else - nr_unpacked++; + for (i = 0; i < 2; i++) { + bch2_bkey_format_add_pos(&format[i], n[i]->data->min_key); + bch2_bkey_format_add_pos(&format[i], n[i]->data->max_key); - prev = k; - k = n; + n[i]->data->format = bch2_bkey_format_done(&format[i]); + btree_node_set_format(n[i], n[i]->data->format); } - BUG_ON(!prev); + u64s = 0; + for_each_btree_node_key(b, k, &iter) { + if (bkey_deleted(k)) + continue; - btree_set_max(n1, bkey_unpack_pos(n1, prev)); - btree_set_min(n2, btree_type_successor(n1->btree_id, n1->key.k.p)); + i = u64s >= n1_u64s; + u64s += k->u64s; - 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)); + if (bch2_bkey_transform(&n[i]->format, out[i], bkey_packed(k) + ? &b->format: &bch2_bkey_format_current, k)) + out[i]->format = KEY_FORMAT_LOCAL_BTREE; + else + bch2_bkey_unpack(b, (void *) out[i], k); - set_btree_bset_end(n1, n1->set); - set_btree_bset_end(n2, n2->set); + out[i]->needs_whiteout = false; - n2->nr.live_u64s = le16_to_cpu(set2->u64s); - n2->nr.bset_u64s[0] = le16_to_cpu(set2->u64s); - n2->nr.packed_keys = n1->nr.packed_keys - nr_packed; - n2->nr.unpacked_keys = n1->nr.unpacked_keys - nr_unpacked; + btree_keys_account_key_add(&n[i]->nr, 0, out[i]); + out[i] = bkey_next(out[i]); + } - n1->nr.live_u64s = le16_to_cpu(set1->u64s); - n1->nr.bset_u64s[0] = le16_to_cpu(set1->u64s); - n1->nr.packed_keys = nr_packed; - n1->nr.unpacked_keys = nr_unpacked; + for (i = 0; i < 2; i++) { + bsets[i]->u64s = cpu_to_le16((u64 *) out[i] - bsets[i]->_data); - BUG_ON(!set1->u64s); - BUG_ON(!set2->u64s); + BUG_ON(!bsets[i]->u64s); - memcpy_u64s(set2->start, - vstruct_end(set1), - le16_to_cpu(set2->u64s)); + set_btree_bset_end(n[i], n[i]->set); - btree_node_reset_sib_u64s(n1); - btree_node_reset_sib_u64s(n2); + btree_node_reset_sib_u64s(n[i]); - bch2_verify_btree_nr_keys(n1); - bch2_verify_btree_nr_keys(n2); + bch2_verify_btree_nr_keys(n[i]); - if (n1->level) { - btree_node_interior_verify(n1); - btree_node_interior_verify(n2); + if (b->c.level) + btree_node_interior_verify(as->c, n[i]); } - - return n2; } /* @@ -1365,86 +1442,73 @@ static struct btree *__btree_split_node(struct btree_update *as, * nodes that were coalesced, and thus in the middle of a child node post * coalescing: */ -static void btree_split_insert_keys(struct btree_update *as, struct btree *b, - struct btree_iter *iter, +static void btree_split_insert_keys(struct btree_update *as, + struct btree_trans *trans, + struct btree_path *path, + struct btree *b, struct keylist *keys) { - struct btree_node_iter node_iter; - struct bkey_i *k = bch2_keylist_front(keys); - struct bkey_packed *src, *dst, *n; - struct bset *i; - - BUG_ON(btree_node_type(b) != BKEY_TYPE_BTREE); - - bch2_btree_node_iter_init(&node_iter, b, &k->k.p); - - while (!bch2_keylist_empty(keys)) { - k = bch2_keylist_front(keys); + if (!bch2_keylist_empty(keys) && + bpos_le(bch2_keylist_front(keys)->k.p, b->data->max_key)) { + struct btree_node_iter node_iter; - 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_btree_node_iter_init(&node_iter, b, &bch2_keylist_front(keys)->k.p); - bch2_insert_fixup_btree_ptr(as, b, iter, k, &node_iter); - bch2_keylist_pop_front(keys); - } + __bch2_btree_insert_keys_interior(as, trans, path, b, node_iter, keys); - /* - * We can't tolerate whiteouts here - with whiteouts there can be - * duplicate keys, and it would be rather bad if we picked a duplicate - * for the pivot: - */ - i = btree_bset_first(b); - src = dst = i->start; - while (src != vstruct_last(i)) { - n = bkey_next_skip_noops(src, vstruct_last(i)); - if (!bkey_deleted(src)) { - memmove_u64s_down(dst, src, src->u64s); - dst = bkey_next(dst); - } - src = n; + btree_node_interior_verify(as->c, b); } - - i->u64s = cpu_to_le16((u64 *) dst - i->_data); - set_btree_bset_end(b, b->set); - - BUG_ON(b->nsets != 1 || - b->nr.live_u64s != le16_to_cpu(btree_bset_first(b)->u64s)); - - btree_node_interior_verify(b); } -static void btree_split(struct btree_update *as, struct btree *b, - struct btree_iter *iter, struct keylist *keys, - unsigned flags) +static int btree_split(struct btree_update *as, struct btree_trans *trans, + struct btree_path *path, struct btree *b, + struct keylist *keys, unsigned flags) { struct bch_fs *c = as->c; - struct btree *parent = btree_node_parent(iter, b); + struct btree *parent = btree_node_parent(path, b); struct btree *n1, *n2 = NULL, *n3 = NULL; + struct btree_path *path1 = NULL, *path2 = NULL; u64 start_time = local_clock(); + int ret = 0; BUG_ON(!parent && (b != btree_node_root(c, b))); - BUG_ON(!btree_node_intent_locked(iter, btree_node_root(c, b)->level)); + BUG_ON(parent && !btree_node_intent_locked(path, b->c.level + 1)); bch2_btree_interior_update_will_free_node(as, b); - n1 = bch2_btree_node_alloc_replacement(as, b); + if (b->nr.live_u64s > BTREE_SPLIT_THRESHOLD(c)) { + struct btree *n[2]; - if (keys) - btree_split_insert_keys(as, n1, iter, keys); + trace_and_count(c, btree_node_split, c, b); - if (bset_u64s(&n1->set[0]) > BTREE_SPLIT_THRESHOLD(c)) { - trace_btree_split(c, b); + n[0] = n1 = bch2_btree_node_alloc(as, trans, b->c.level); + n[1] = n2 = bch2_btree_node_alloc(as, trans, b->c.level); - n2 = __btree_split_node(as, n1, iter); + __btree_split_node(as, trans, b, n); + + if (keys) { + btree_split_insert_keys(as, trans, path, n1, keys); + btree_split_insert_keys(as, trans, path, n2, keys); + BUG_ON(!bch2_keylist_empty(keys)); + } bch2_btree_build_aux_trees(n2); bch2_btree_build_aux_trees(n1); - six_unlock_write(&n2->lock); - six_unlock_write(&n1->lock); - bch2_btree_node_write(c, n2, SIX_LOCK_intent); + bch2_btree_update_add_new_node(as, n1); + bch2_btree_update_add_new_node(as, n2); + six_unlock_write(&n2->c.lock); + six_unlock_write(&n1->c.lock); + + path1 = get_unlocked_mut_path(trans, path->btree_id, n1->c.level, n1->key.k.p); + six_lock_increment(&n1->c.lock, SIX_LOCK_intent); + mark_btree_node_locked(trans, path1, n1->c.level, SIX_LOCK_intent); + bch2_btree_path_level_init(trans, path1, n1); + + path2 = get_unlocked_mut_path(trans, path->btree_id, n2->c.level, n2->key.k.p); + six_lock_increment(&n2->c.lock, SIX_LOCK_intent); + mark_btree_node_locked(trans, path2, n2->c.level, SIX_LOCK_intent); + bch2_btree_path_level_init(trans, path2, n2); /* * Note that on recursive parent_keys == keys, so we @@ -1456,53 +1520,69 @@ 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, trans, b->c.level + 1); + + bch2_btree_update_add_new_node(as, n3); + six_unlock_write(&n3->c.lock); + + path2->locks_want++; + BUG_ON(btree_node_locked(path2, n3->c.level)); + six_lock_increment(&n3->c.lock, SIX_LOCK_intent); + mark_btree_node_locked(trans, path2, n3->c.level, SIX_LOCK_intent); + bch2_btree_path_level_init(trans, path2, n3); n3->sib_u64s[0] = U16_MAX; n3->sib_u64s[1] = U16_MAX; - btree_split_insert_keys(as, n3, iter, &as->parent_keys); - - bch2_btree_node_write(c, n3, SIX_LOCK_intent); + btree_split_insert_keys(as, trans, path, n3, &as->parent_keys); } } else { - trace_btree_compact(c, b); + trace_and_count(c, btree_node_compact, c, b); + + n1 = bch2_btree_node_alloc_replacement(as, trans, b); + + if (keys) { + btree_split_insert_keys(as, trans, path, n1, keys); + BUG_ON(!bch2_keylist_empty(keys)); + } bch2_btree_build_aux_trees(n1); - six_unlock_write(&n1->lock); + bch2_btree_update_add_new_node(as, n1); + six_unlock_write(&n1->c.lock); - bch2_keylist_add(&as->parent_keys, &n1->key); - } + path1 = get_unlocked_mut_path(trans, path->btree_id, n1->c.level, n1->key.k.p); + six_lock_increment(&n1->c.lock, SIX_LOCK_intent); + mark_btree_node_locked(trans, path1, n1->c.level, SIX_LOCK_intent); + bch2_btree_path_level_init(trans, path1, n1); - bch2_btree_node_write(c, n1, SIX_LOCK_intent); + if (parent) + bch2_keylist_add(&as->parent_keys, &n1->key); + } /* New nodes all written, now make them visible: */ if (parent) { /* Split a non root node */ - bch2_btree_insert_node(as, parent, iter, &as->parent_keys, flags); + ret = bch2_btree_insert_node(as, trans, path, parent, &as->parent_keys, flags); + if (ret) + goto err; } else if (n3) { - bch2_btree_set_root(as, n3, iter); + bch2_btree_set_root(as, trans, path, n3); } else { /* Root filled up but didn't need to be split */ - bch2_btree_set_root(as, n1, iter); + bch2_btree_set_root(as, trans, path, n1); } - bch2_open_buckets_put(c, &n1->ob); - if (n2) - bch2_open_buckets_put(c, &n2->ob); - if (n3) - bch2_open_buckets_put(c, &n3->ob); - - /* Successful split, update the iterator to point to the new nodes: */ - - six_lock_increment(&b->lock, SIX_LOCK_intent); - bch2_btree_iter_node_drop(iter, b); - if (n3) - bch2_btree_iter_node_replace(iter, n3); - if (n2) - bch2_btree_iter_node_replace(iter, n2); - bch2_btree_iter_node_replace(iter, n1); + if (n3) { + bch2_btree_update_get_open_buckets(as, n3); + bch2_btree_node_write(c, n3, SIX_LOCK_intent, 0); + } + if (n2) { + bch2_btree_update_get_open_buckets(as, n2); + bch2_btree_node_write(c, n2, SIX_LOCK_intent, 0); + } + bch2_btree_update_get_open_buckets(as, n1); + bch2_btree_node_write(c, n1, SIX_LOCK_intent, 0); /* * The old node must be freed (in memory) _before_ unlocking the new @@ -1510,54 +1590,63 @@ static void btree_split(struct btree_update *as, struct btree *b, * node after another thread has locked and updated the new node, thus * seeing stale data: */ - bch2_btree_node_free_inmem(c, b, iter); + bch2_btree_node_free_inmem(trans, path, b); + + if (n3) + bch2_trans_node_add(trans, n3); + if (n2) + bch2_trans_node_add(trans, n2); + bch2_trans_node_add(trans, n1); 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); +out: + if (path2) { + __bch2_btree_path_unlock(trans, path2); + bch2_path_put(trans, path2, true); + } + if (path1) { + __bch2_btree_path_unlock(trans, path1); + bch2_path_put(trans, path1, true); + } - bch2_btree_trans_verify_locks(iter->trans); + bch2_trans_verify_locks(trans); - bch2_time_stats_update(&c->times[BCH_TIME_btree_node_split], + bch2_time_stats_update(&c->times[n2 + ? BCH_TIME_btree_node_split + : BCH_TIME_btree_node_compact], start_time); + return ret; +err: + if (n3) + bch2_btree_node_free_never_used(as, trans, n3); + if (n2) + bch2_btree_node_free_never_used(as, trans, n2); + bch2_btree_node_free_never_used(as, trans, n1); + goto out; } static void -bch2_btree_insert_keys_interior(struct btree_update *as, struct btree *b, - struct btree_iter *iter, struct keylist *keys) +bch2_btree_insert_keys_interior(struct btree_update *as, + struct btree_trans *trans, + struct btree_path *path, + struct btree *b, + struct keylist *keys) { - struct btree_iter *linked; - struct btree_node_iter node_iter; - struct bkey_i *insert = bch2_keylist_front(keys); - struct bkey_packed *k; + struct btree_path *linked; - /* Don't screw up @iter's position: */ - node_iter = iter->l[b->level].iter; + __bch2_btree_insert_keys_interior(as, trans, path, b, + path->l[b->c.level].iter, keys); - /* - * btree_split(), btree_gc_coalesce() will insert keys before - * the iterator's current position - they know the keys go in - * the node the iterator points to: - */ - while ((k = bch2_btree_node_iter_prev_all(&node_iter, b)) && - (bkey_cmp_packed(b, k, &insert->k) >= 0)) - ; + btree_update_updated_node(as, b); - while (!bch2_keylist_empty(keys)) { - insert = bch2_keylist_front(keys); + trans_for_each_path_with_node(trans, b, linked) + bch2_btree_node_iter_peek(&linked->l[b->c.level].iter, b); - 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_trans_verify_iters(iter->trans, b); + bch2_trans_verify_paths(trans); } /** @@ -1572,31 +1661,39 @@ bch2_btree_insert_keys_interior(struct btree_update *as, struct btree *b, * If a split occurred, this function will return early. This can only happen * for leaf nodes -- inserts into interior nodes have to be atomic. */ -void bch2_btree_insert_node(struct btree_update *as, struct btree *b, - struct btree_iter *iter, struct keylist *keys, - unsigned flags) +static int bch2_btree_insert_node(struct btree_update *as, struct btree_trans *trans, + struct btree_path *path, struct btree *b, + struct keylist *keys, unsigned flags) { struct bch_fs *c = as->c; int old_u64s = le16_to_cpu(btree_bset_last(b)->u64s); int old_live_u64s = b->nr.live_u64s; int live_u64s_added, u64s_added; + int ret; - BUG_ON(!btree_node_intent_locked(iter, btree_node_root(c, b)->level)); - BUG_ON(!b->level); + lockdep_assert_held(&c->gc_lock); + BUG_ON(!btree_node_intent_locked(path, b->c.level)); + BUG_ON(!b->c.level); BUG_ON(!as || as->b); bch2_verify_keylist_sorted(keys); - if (as->must_rewrite) - goto split; + if (!(local_clock() & 63)) + return btree_trans_restart(trans, BCH_ERR_transaction_restart_split_race); - bch2_btree_node_lock_for_insert(c, b, iter); + ret = bch2_btree_node_lock_write(trans, path, &b->c); + if (ret) + return ret; - if (!bch2_btree_node_insert_fits(c, b, bch_keylist_u64s(keys))) { - bch2_btree_node_unlock_write(b, iter); + bch2_btree_node_prep_for_write(trans, path, b); + + if (!bch2_btree_node_insert_fits(c, b, bch2_keylist_u64s(keys))) { + bch2_btree_node_unlock_write(trans, path, b); goto split; } - bch2_btree_insert_keys_interior(as, b, iter, keys); + btree_node_interior_verify(c, b); + + bch2_btree_insert_keys_interior(as, trans, path, b, keys); live_u64s_added = (int) b->nr.live_u64s - old_live_u64s; u64s_added = (int) le16_to_cpu(btree_bset_last(b)->u64s) - old_u64s; @@ -1608,135 +1705,98 @@ void bch2_btree_insert_node(struct btree_update *as, struct btree *b, if (u64s_added > live_u64s_added && bch2_maybe_compact_whiteouts(c, b)) - bch2_btree_iter_reinit_node(iter, b); - - bch2_btree_node_unlock_write(b, iter); + bch2_trans_node_reinit_iter(trans, b); - btree_node_interior_verify(b); + bch2_btree_node_unlock_write(trans, path, b); + btree_node_interior_verify(c, b); + return 0; +split: /* - * when called from the btree_split path the new nodes aren't added to - * the btree iterator yet, so the merge path's unlock/wait/relock dance - * won't work: + * We could attempt to avoid the transaction restart, by calling + * bch2_btree_path_upgrade() and allocating more nodes: */ - bch2_foreground_maybe_merge(c, iter, b->level, - flags|BTREE_INSERT_NOUNLOCK); - return; -split: - btree_split(as, b, iter, keys, flags); + if (b->c.level >= as->update_level) { + trace_and_count(c, trans_restart_split_race, trans, _THIS_IP_); + return btree_trans_restart(trans, BCH_ERR_transaction_restart_split_race); + } + + return btree_split(as, trans, path, b, keys, flags); } -int bch2_btree_split_leaf(struct bch_fs *c, struct btree_iter *iter, +int bch2_btree_split_leaf(struct btree_trans *trans, + struct btree_path *path, unsigned flags) { - struct btree_trans *trans = iter->trans; - struct btree *b = iter->l[0].b; + struct btree *b = path_l(path)->b; struct btree_update *as; - struct closure cl; + unsigned l; int ret = 0; - struct btree_iter *linked; - /* - * We already have a disk reservation and open buckets pinned; this - * allocation must not block: - */ - trans_for_each_iter(trans, linked) - if (linked->btree_id == BTREE_ID_EXTENTS) - flags |= BTREE_INSERT_USE_RESERVE; - - closure_init_stack(&cl); - - /* 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) - return -EINTR; - - bch2_trans_unlock(trans); - down_read(&c->gc_lock); + as = bch2_btree_update_start(trans, path, path->level, + true, flags); + if (IS_ERR(as)) + return PTR_ERR(as); - if (!bch2_trans_relock(trans)) - ret = -EINTR; + ret = btree_split(as, trans, path, b, NULL, flags); + if (ret) { + bch2_btree_update_free(as, trans); + return ret; } - /* - * XXX: figure out how far we might need to split, - * instead of locking/reserving all the way to the root: - */ - if (!bch2_btree_iter_upgrade(iter, U8_MAX)) { - trace_trans_restart_iter_upgrade(trans->ip); - ret = -EINTR; - goto out; - } + bch2_btree_update_done(as, trans); - as = bch2_btree_update_start(c, iter->btree_id, - btree_update_reserve_required(c, b), flags, - !(flags & BTREE_INSERT_NOUNLOCK) ? &cl : NULL); - if (IS_ERR(as)) { - ret = PTR_ERR(as); - if (ret == -EAGAIN) { - BUG_ON(flags & BTREE_INSERT_NOUNLOCK); - bch2_trans_unlock(trans); - ret = -EINTR; - } - goto out; - } + for (l = path->level + 1; btree_node_intent_locked(path, l) && !ret; l++) + ret = bch2_foreground_maybe_merge(trans, path, l, flags); - btree_split(as, b, iter, NULL, flags); - bch2_btree_update_done(as); - - /* - * We haven't successfully inserted yet, so don't downgrade all the way - * back to read locks; - */ - __bch2_btree_iter_downgrade(iter, 1); -out: - if (!(flags & BTREE_INSERT_GC_LOCK_HELD)) - up_read(&c->gc_lock); - closure_sync(&cl); return ret; } -void __bch2_foreground_maybe_merge(struct bch_fs *c, - struct btree_iter *iter, - unsigned level, - unsigned flags, - enum btree_node_sibling sib) +int __bch2_foreground_maybe_merge(struct btree_trans *trans, + struct btree_path *path, + unsigned level, + unsigned flags, + enum btree_node_sibling sib) { - struct btree_trans *trans = iter->trans; + struct bch_fs *c = trans->c; + struct btree_path *sib_path = NULL, *new_path = NULL; struct btree_update *as; struct bkey_format_state new_s; struct bkey_format new_f; struct bkey_i delete; struct btree *b, *m, *n, *prev, *next, *parent; - struct closure cl; + struct bpos sib_pos; size_t sib_u64s; + u64 start_time = local_clock(); int ret = 0; - BUG_ON(!btree_node_locked(iter, level)); - - closure_init_stack(&cl); -retry: - BUG_ON(!btree_node_locked(iter, level)); + BUG_ON(!path->should_be_locked); + BUG_ON(!btree_node_locked(path, level)); - b = iter->l[level].b; + b = path->l[level].b; - parent = btree_node_parent(iter, b); - if (!parent) - goto out; + if ((sib == btree_prev_sib && bpos_eq(b->data->min_key, POS_MIN)) || + (sib == btree_next_sib && bpos_eq(b->data->max_key, SPOS_MAX))) { + b->sib_u64s[sib] = U16_MAX; + return 0; + } - if (b->sib_u64s[sib] > BTREE_FOREGROUND_MERGE_THRESHOLD(c)) - goto out; + sib_pos = sib == btree_prev_sib + ? bpos_predecessor(b->data->min_key) + : bpos_successor(b->data->max_key); - /* XXX: can't be holding read locks */ - m = bch2_btree_node_get_sibling(c, iter, b, sib); - if (IS_ERR(m)) { - ret = PTR_ERR(m); + sib_path = bch2_path_get(trans, path->btree_id, sib_pos, + U8_MAX, level, BTREE_ITER_INTENT, _THIS_IP_); + ret = bch2_btree_path_traverse(trans, sib_path, false); + if (ret) goto err; - } - /* NULL means no sibling: */ - if (!m) { + btree_path_set_should_be_locked(sib_path); + + m = sib_path->l[level].b; + + if (btree_node_parent(path, b) != + btree_node_parent(sib_path, m)) { b->sib_u64s[sib] = U16_MAX; goto out; } @@ -1749,9 +1809,28 @@ retry: next = m; } + if (!bpos_eq(bpos_successor(prev->data->max_key), next->data->min_key)) { + struct printbuf buf1 = PRINTBUF, buf2 = PRINTBUF; + + bch2_bpos_to_text(&buf1, prev->data->max_key); + bch2_bpos_to_text(&buf2, next->data->min_key); + bch_err(c, + "%s(): btree topology error:\n" + " prev ends at %s\n" + " next starts at %s", + __func__, buf1.buf, buf2.buf); + printbuf_exit(&buf1); + printbuf_exit(&buf2); + bch2_topology_error(c); + ret = -EIO; + goto err; + } + bch2_bkey_format_init(&new_s); - __bch2_btree_calc_format(&new_s, b); - __bch2_btree_calc_format(&new_s, m); + bch2_bkey_format_add_pos(&new_s, prev->data->min_key); + __bch2_btree_calc_format(&new_s, prev); + __bch2_btree_calc_format(&new_s, next); + bch2_bkey_format_add_pos(&new_s, next->data->max_key); new_f = bch2_bkey_format_done(&new_s); sib_u64s = btree_node_u64s_with_format(b, &new_f) + @@ -1764,449 +1843,471 @@ retry: } sib_u64s = min(sib_u64s, btree_max_u64s(c)); + sib_u64s = min(sib_u64s, (size_t) U16_MAX - 1); b->sib_u64s[sib] = sib_u64s; - if (b->sib_u64s[sib] > BTREE_FOREGROUND_MERGE_THRESHOLD(c)) { - six_unlock_intent(&m->lock); + if (b->sib_u64s[sib] > c->btree_foreground_merge_threshold) goto out; - } - /* We're changing btree topology, doesn't mix with gc: */ - if (!(flags & BTREE_INSERT_GC_LOCK_HELD) && - !down_read_trylock(&c->gc_lock)) - goto err_cycle_gc_lock; - - if (!bch2_btree_iter_upgrade(iter, U8_MAX)) { - ret = -EINTR; - goto err_unlock; - } - - as = bch2_btree_update_start(c, iter->btree_id, - btree_update_reserve_required(c, parent) + 1, + parent = btree_node_parent(path, b); + as = bch2_btree_update_start(trans, path, level, false, BTREE_INSERT_NOFAIL| - BTREE_INSERT_USE_RESERVE, - !(flags & BTREE_INSERT_NOUNLOCK) ? &cl : NULL); - if (IS_ERR(as)) { - ret = PTR_ERR(as); - goto err_unlock; - } + BTREE_INSERT_USE_RESERVE| + flags); + ret = PTR_ERR_OR_ZERO(as); + if (ret) + goto err; - trace_btree_merge(c, b); + trace_and_count(c, btree_node_merge, c, b); 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, trans, b->c.level); + + SET_BTREE_NODE_SEQ(n->data, + max(BTREE_NODE_SEQ(b->data), + BTREE_NODE_SEQ(m->data)) + 1); btree_set_min(n, prev->data->min_key); btree_set_max(n, next->data->max_key); - n->data->format = new_f; + n->data->format = new_f; btree_node_set_format(n, new_f); bch2_btree_sort_into(c, n, prev); bch2_btree_sort_into(c, n, next); bch2_btree_build_aux_trees(n); - six_unlock_write(&n->lock); + bch2_btree_update_add_new_node(as, n); + six_unlock_write(&n->c.lock); + + new_path = get_unlocked_mut_path(trans, path->btree_id, n->c.level, n->key.k.p); + six_lock_increment(&n->c.lock, SIX_LOCK_intent); + mark_btree_node_locked(trans, new_path, n->c.level, SIX_LOCK_intent); + bch2_btree_path_level_init(trans, new_path, n); bkey_init(&delete.k); delete.k.p = prev->key.k.p; bch2_keylist_add(&as->parent_keys, &delete); bch2_keylist_add(&as->parent_keys, &n->key); - bch2_btree_node_write(c, n, SIX_LOCK_intent); + bch2_trans_verify_paths(trans); - bch2_btree_insert_node(as, parent, iter, &as->parent_keys, flags); + ret = bch2_btree_insert_node(as, trans, path, parent, &as->parent_keys, flags); + if (ret) + goto err_free_update; - bch2_open_buckets_put(c, &n->ob); + bch2_trans_verify_paths(trans); - six_lock_increment(&b->lock, SIX_LOCK_intent); - bch2_btree_iter_node_drop(iter, b); - bch2_btree_iter_node_drop(iter, m); + bch2_btree_update_get_open_buckets(as, n); + bch2_btree_node_write(c, n, SIX_LOCK_intent, 0); - bch2_btree_iter_node_replace(iter, n); + bch2_btree_node_free_inmem(trans, path, b); + bch2_btree_node_free_inmem(trans, sib_path, m); - bch2_btree_trans_verify_iters(trans, n); + bch2_trans_node_add(trans, n); - bch2_btree_node_free_inmem(c, b, iter); - bch2_btree_node_free_inmem(c, m, iter); + bch2_trans_verify_paths(trans); - six_unlock_intent(&n->lock); + six_unlock_intent(&n->c.lock); - bch2_btree_update_done(as); + bch2_btree_update_done(as, trans); - if (!(flags & BTREE_INSERT_GC_LOCK_HELD)) - up_read(&c->gc_lock); + bch2_time_stats_update(&c->times[BCH_TIME_btree_node_merge], start_time); out: - bch2_btree_trans_verify_locks(trans); - - /* - * Don't downgrade locks here: we're called after successful insert, - * and the caller will downgrade locks after a successful insert - * anyways (in case e.g. a split was required first) - * - * And we're also called when inserting into interior nodes in the - * split path, and downgrading to read locks in there is potentially - * confusing: - */ - closure_sync(&cl); - return; - -err_cycle_gc_lock: - six_unlock_intent(&m->lock); - - if (flags & BTREE_INSERT_NOUNLOCK) - goto out; - - bch2_trans_unlock(trans); - - down_read(&c->gc_lock); - up_read(&c->gc_lock); - ret = -EINTR; - goto err; - -err_unlock: - six_unlock_intent(&m->lock); - if (!(flags & BTREE_INSERT_GC_LOCK_HELD)) - up_read(&c->gc_lock); err: - BUG_ON(ret == -EAGAIN && (flags & BTREE_INSERT_NOUNLOCK)); - - if ((ret == -EAGAIN || ret == -EINTR) && - !(flags & BTREE_INSERT_NOUNLOCK)) { - bch2_trans_unlock(trans); - closure_sync(&cl); - ret = bch2_btree_iter_traverse(iter); - if (ret) - goto out; - - goto retry; - } - + if (new_path) + bch2_path_put(trans, new_path, true); + bch2_path_put(trans, sib_path, true); + bch2_trans_verify_locks(trans); + return ret; +err_free_update: + bch2_btree_node_free_never_used(as, trans, n); + bch2_btree_update_free(as, trans); goto out; } -static int __btree_node_rewrite(struct bch_fs *c, struct btree_iter *iter, - struct btree *b, unsigned flags, - struct closure *cl) +/** + * bch_btree_node_rewrite - Rewrite/move a btree node + */ +int bch2_btree_node_rewrite(struct btree_trans *trans, + struct btree_iter *iter, + struct btree *b, + unsigned flags) { - struct btree *n, *parent = btree_node_parent(iter, b); + struct bch_fs *c = trans->c; + struct btree_path *new_path = NULL; + struct btree *n, *parent; struct btree_update *as; + int ret; - as = bch2_btree_update_start(c, iter->btree_id, - (parent - ? btree_update_reserve_required(c, parent) - : 0) + 1, - flags, cl); - if (IS_ERR(as)) { - trace_btree_gc_rewrite_node_fail(c, b); - return PTR_ERR(as); - } + flags |= BTREE_INSERT_NOFAIL; + + parent = btree_node_parent(iter->path, b); + as = bch2_btree_update_start(trans, iter->path, b->c.level, + false, flags); + ret = PTR_ERR_OR_ZERO(as); + if (ret) + goto out; bch2_btree_interior_update_will_free_node(as, b); - n = bch2_btree_node_alloc_replacement(as, b); + n = bch2_btree_node_alloc_replacement(as, trans, b); bch2_btree_build_aux_trees(n); - six_unlock_write(&n->lock); + bch2_btree_update_add_new_node(as, n); + six_unlock_write(&n->c.lock); - trace_btree_gc_rewrite_node(c, b); + new_path = get_unlocked_mut_path(trans, iter->btree_id, n->c.level, n->key.k.p); + six_lock_increment(&n->c.lock, SIX_LOCK_intent); + mark_btree_node_locked(trans, new_path, n->c.level, SIX_LOCK_intent); + bch2_btree_path_level_init(trans, new_path, n); - bch2_btree_node_write(c, n, SIX_LOCK_intent); + trace_and_count(c, btree_node_rewrite, c, b); if (parent) { bch2_keylist_add(&as->parent_keys, &n->key); - bch2_btree_insert_node(as, parent, iter, &as->parent_keys, flags); + ret = bch2_btree_insert_node(as, trans, iter->path, parent, + &as->parent_keys, flags); + if (ret) + goto err; } else { - bch2_btree_set_root(as, n, iter); + bch2_btree_set_root(as, trans, iter->path, n); } - bch2_open_buckets_put(c, &n->ob); + bch2_btree_update_get_open_buckets(as, n); + bch2_btree_node_write(c, n, SIX_LOCK_intent, 0); - six_lock_increment(&b->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); + bch2_btree_node_free_inmem(trans, iter->path, b); - bch2_btree_update_done(as); - return 0; + bch2_trans_node_add(trans, n); + six_unlock_intent(&n->c.lock); + + bch2_btree_update_done(as, trans); +out: + if (new_path) + bch2_path_put(trans, new_path, true); + bch2_btree_path_downgrade(trans, iter->path); + return ret; +err: + bch2_btree_node_free_never_used(as, trans, n); + bch2_btree_update_free(as, trans); + goto out; } -/** - * bch_btree_node_rewrite - Rewrite/move a btree node - * - * Returns 0 on success, -EINTR or -EAGAIN on failure (i.e. - * btree_check_reserve() has to wait) - */ -int bch2_btree_node_rewrite(struct bch_fs *c, struct btree_iter *iter, - __le64 seq, unsigned flags) +struct async_btree_rewrite { + struct bch_fs *c; + struct work_struct work; + struct list_head list; + enum btree_id btree_id; + unsigned level; + struct bpos pos; + __le64 seq; +}; + +static int async_btree_node_rewrite_trans(struct btree_trans *trans, + struct async_btree_rewrite *a) { - struct btree_trans *trans = iter->trans; - struct closure cl; + struct bch_fs *c = trans->c; + struct btree_iter iter; struct btree *b; int ret; - flags |= BTREE_INSERT_NOFAIL; + bch2_trans_node_iter_init(trans, &iter, a->btree_id, a->pos, + BTREE_MAX_DEPTH, a->level, 0); + b = bch2_btree_iter_peek_node(&iter); + ret = PTR_ERR_OR_ZERO(b); + if (ret) + goto out; - closure_init_stack(&cl); + if (!b || b->data->keys.seq != a->seq) { + struct printbuf buf = PRINTBUF; - bch2_btree_iter_upgrade(iter, U8_MAX); + if (b) + bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(&b->key)); + else + prt_str(&buf, "(null"); + bch_info(c, "%s: node to rewrite not found:, searching for seq %llu, got\n%s", + __func__, a->seq, buf.buf); + printbuf_exit(&buf); + goto out; + } - if (!(flags & BTREE_INSERT_GC_LOCK_HELD)) { - if (!down_read_trylock(&c->gc_lock)) { - bch2_trans_unlock(trans); - down_read(&c->gc_lock); + ret = bch2_btree_node_rewrite(trans, &iter, b, 0); +out: + bch2_trans_iter_exit(trans, &iter); + + return ret; +} + +void async_btree_node_rewrite_work(struct work_struct *work) +{ + struct async_btree_rewrite *a = + container_of(work, struct async_btree_rewrite, work); + struct bch_fs *c = a->c; + int ret; + + ret = bch2_trans_do(c, NULL, NULL, 0, + async_btree_node_rewrite_trans(&trans, a)); + if (ret) + bch_err(c, "%s: error %s", __func__, bch2_err_str(ret)); + bch2_write_ref_put(c, BCH_WRITE_REF_node_rewrite); + kfree(a); +} + +void bch2_btree_node_rewrite_async(struct bch_fs *c, struct btree *b) +{ + struct async_btree_rewrite *a; + int ret; + + a = kmalloc(sizeof(*a), GFP_NOFS); + if (!a) { + bch_err(c, "%s: error allocating memory", __func__); + return; + } + + a->c = c; + a->btree_id = b->c.btree_id; + a->level = b->c.level; + a->pos = b->key.k.p; + a->seq = b->data->keys.seq; + INIT_WORK(&a->work, async_btree_node_rewrite_work); + + if (unlikely(!test_bit(BCH_FS_MAY_GO_RW, &c->flags))) { + mutex_lock(&c->pending_node_rewrites_lock); + list_add(&a->list, &c->pending_node_rewrites); + mutex_unlock(&c->pending_node_rewrites_lock); + return; + } + + if (!bch2_write_ref_tryget(c, BCH_WRITE_REF_node_rewrite)) { + if (test_bit(BCH_FS_STARTED, &c->flags)) { + bch_err(c, "%s: error getting c->writes ref", __func__); + kfree(a); + return; } + + ret = bch2_fs_read_write_early(c); + if (ret) { + bch_err(c, "%s: error going read-write: %s", + __func__, bch2_err_str(ret)); + kfree(a); + return; + } + + bch2_write_ref_get(c, BCH_WRITE_REF_node_rewrite); } - while (1) { - ret = bch2_btree_iter_traverse(iter); - if (ret) - break; + queue_work(c->btree_interior_update_worker, &a->work); +} - b = bch2_btree_iter_peek_node(iter); - if (!b || b->data->keys.seq != seq) - break; +void bch2_do_pending_node_rewrites(struct bch_fs *c) +{ + struct async_btree_rewrite *a, *n; - ret = __btree_node_rewrite(c, iter, b, flags, &cl); - if (ret != -EAGAIN && - ret != -EINTR) - break; + mutex_lock(&c->pending_node_rewrites_lock); + list_for_each_entry_safe(a, n, &c->pending_node_rewrites, list) { + list_del(&a->list); - bch2_trans_unlock(trans); - closure_sync(&cl); + bch2_write_ref_get(c, BCH_WRITE_REF_node_rewrite); + queue_work(c->btree_interior_update_worker, &a->work); } + mutex_unlock(&c->pending_node_rewrites_lock); +} - bch2_btree_iter_downgrade(iter); +void bch2_free_pending_node_rewrites(struct bch_fs *c) +{ + struct async_btree_rewrite *a, *n; - if (!(flags & BTREE_INSERT_GC_LOCK_HELD)) - up_read(&c->gc_lock); + mutex_lock(&c->pending_node_rewrites_lock); + list_for_each_entry_safe(a, n, &c->pending_node_rewrites, list) { + list_del(&a->list); - closure_sync(&cl); - return ret; + kfree(a); + } + mutex_unlock(&c->pending_node_rewrites_lock); } -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 *new_key) +static int __bch2_btree_node_update_key(struct btree_trans *trans, + struct btree_iter *iter, + struct btree *b, struct btree *new_hash, + struct bkey_i *new_key, + bool skip_triggers) { + struct bch_fs *c = trans->c; + struct btree_iter iter2 = { NULL }; 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; + if (!skip_triggers) { + ret = bch2_trans_mark_old(trans, b->c.btree_id, b->c.level + 1, + bkey_i_to_s_c(&b->key), 0); + if (ret) + return ret; - btree_interior_update_add_node_reference(as, b); + ret = bch2_trans_mark_new(trans, b->c.btree_id, b->c.level + 1, + new_key, 0); + if (ret) + return ret; + } - /* - * 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)), - BCH_DISK_RESERVATION_NOFAIL); - BUG_ON(ret); + if (new_hash) { + bkey_copy(&new_hash->key, new_key); + ret = bch2_btree_node_hash_insert(&c->btree_cache, + new_hash, b->c.level, b->c.btree_id); + BUG_ON(ret); + } - parent = btree_node_parent(iter, b); + parent = btree_node_parent(iter->path, b); if (parent) { - if (new_hash) { - bkey_copy(&new_hash->key, new_key); - ret = bch2_btree_node_hash_insert(&c->btree_cache, - new_hash, b->level, b->btree_id); - BUG_ON(ret); - } + bch2_trans_copy_iter(&iter2, iter); - bch2_keylist_add(&as->parent_keys, new_key); - bch2_btree_insert_node(as, parent, iter, &as->parent_keys, 0); + iter2.path = bch2_btree_path_make_mut(trans, iter2.path, + iter2.flags & BTREE_ITER_INTENT, + _THIS_IP_); - if (new_hash) { - mutex_lock(&c->btree_cache.lock); - bch2_btree_node_hash_remove(&c->btree_cache, new_hash); + BUG_ON(iter2.path->level != b->c.level); + BUG_ON(!bpos_eq(iter2.path->pos, new_key->k.p)); - bch2_btree_node_hash_remove(&c->btree_cache, b); + btree_path_set_level_up(trans, iter2.path); - 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); - } - } else { - struct bch_fs_usage *fs_usage; + trans->paths_sorted = false; + ret = bch2_btree_iter_traverse(&iter2) ?: + bch2_trans_update(trans, &iter2, new_key, BTREE_TRIGGER_NORUN); + if (ret) + goto err; + } else { BUG_ON(btree_node_root(c, b) != b); - bch2_btree_node_lock_write(b, iter); + ret = darray_make_room(&trans->extra_journal_entries, + jset_u64s(new_key->k.u64s)); + if (ret) + return ret; - 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), - 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), - 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); + journal_entry_set((void *) &darray_top(trans->extra_journal_entries), + BCH_JSET_ENTRY_btree_root, + b->c.btree_id, b->c.level, + new_key, new_key->k.u64s); + trans->extra_journal_entries.nr += jset_u64s(new_key->k.u64s); + } - if (btree_ptr_hash_val(new_key) != b->hash_val) { - mutex_lock(&c->btree_cache.lock); - bch2_btree_node_hash_remove(&c->btree_cache, b); + ret = bch2_trans_commit(trans, NULL, NULL, + BTREE_INSERT_NOFAIL| + BTREE_INSERT_NOCHECK_RW| + BTREE_INSERT_USE_RESERVE| + BTREE_INSERT_JOURNAL_RECLAIM| + JOURNAL_WATERMARK_reserved); + if (ret) + goto err; - 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); - } + bch2_btree_node_lock_write_nofail(trans, iter->path, &b->c); - btree_update_updated_root(as); - bch2_btree_node_unlock_write(b, iter); + if (new_hash) { + mutex_lock(&c->btree_cache.lock); + bch2_btree_node_hash_remove(&c->btree_cache, new_hash); + bch2_btree_node_hash_remove(&c->btree_cache, b); + + 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); } - bch2_btree_update_done(as); + bch2_btree_node_unlock_write(trans, iter->path, b); +out: + bch2_trans_iter_exit(trans, &iter2); + return ret; +err: + if (new_hash) { + mutex_lock(&c->btree_cache.lock); + bch2_btree_node_hash_remove(&c->btree_cache, b); + mutex_unlock(&c->btree_cache.lock); + } + goto out; } -int bch2_btree_node_update_key(struct bch_fs *c, struct btree_iter *iter, - struct btree *b, - struct bkey_i *new_key) +int bch2_btree_node_update_key(struct btree_trans *trans, struct btree_iter *iter, + struct btree *b, struct bkey_i *new_key, + bool skip_triggers) { - struct btree *parent = btree_node_parent(iter, b); - struct btree_update *as = NULL; + struct bch_fs *c = trans->c; struct btree *new_hash = NULL; + struct btree_path *path = iter->path; struct closure cl; - int ret; - - closure_init_stack(&cl); - - if (!bch2_btree_iter_upgrade(iter, U8_MAX)) - return -EINTR; + int ret = 0; - if (!down_read_trylock(&c->gc_lock)) { - bch2_trans_unlock(iter->trans); - down_read(&c->gc_lock); + ret = bch2_btree_path_upgrade(trans, path, b->c.level + 1); + if (ret) + return ret; - if (!bch2_trans_relock(iter->trans)) { - ret = -EINTR; - goto err; - } - } + closure_init_stack(&cl); /* * 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) { - bch2_trans_unlock(iter->trans); - up_read(&c->gc_lock); + bch2_trans_unlock(trans); closure_sync(&cl); - down_read(&c->gc_lock); - - if (!bch2_trans_relock(iter->trans)) { - ret = -EINTR; - goto err; - } + ret = bch2_trans_relock(trans); + if (ret) + return ret; } - new_hash = bch2_btree_node_mem_alloc(c); - } - - as = bch2_btree_update_start(c, iter->btree_id, - parent ? btree_update_reserve_required(c, parent) : 0, - BTREE_INSERT_NOFAIL| - BTREE_INSERT_USE_RESERVE| - BTREE_INSERT_USE_ALLOC_RESERVE, - &cl); - - if (IS_ERR(as)) { - ret = PTR_ERR(as); - if (ret == -EAGAIN) - ret = -EINTR; - - if (ret != -EINTR) - goto err; - - bch2_trans_unlock(iter->trans); - up_read(&c->gc_lock); - closure_sync(&cl); - down_read(&c->gc_lock); - - if (!bch2_trans_relock(iter->trans)) - goto err; + new_hash = bch2_btree_node_mem_alloc(c, false); } - ret = bch2_mark_bkey_replicas(c, bkey_i_to_s_c(new_key)); - if (ret) - goto err_free_update; - - __bch2_btree_node_update_key(c, as, iter, b, new_hash, new_key); + path->intent_ref++; + ret = __bch2_btree_node_update_key(trans, iter, b, new_hash, + new_key, skip_triggers); + --path->intent_ref; - bch2_btree_iter_downgrade(iter); -err: if (new_hash) { mutex_lock(&c->btree_cache.lock); 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); + bch2_btree_cache_cannibalize_unlock(c); + return ret; +} + +int bch2_btree_node_update_key_get_iter(struct btree_trans *trans, + struct btree *b, struct bkey_i *new_key, + bool skip_triggers) +{ + struct btree_iter iter; + int ret; + + bch2_trans_node_iter_init(trans, &iter, b->c.btree_id, b->key.k.p, + BTREE_MAX_DEPTH, b->c.level, + BTREE_ITER_INTENT); + ret = bch2_btree_iter_traverse(&iter); + if (ret) + goto out; + + /* has node been freed? */ + if (iter.path->l[b->c.level].b != b) { + /* node has been freed: */ + BUG_ON(!btree_node_dying(b)); + goto out; + } + + BUG_ON(!btree_node_hashed(b)); + + ret = bch2_btree_node_update_key(trans, &iter, b, new_key, skip_triggers); +out: + bch2_trans_iter_exit(trans, &iter); return ret; -err_free_update: - bch2_btree_update_free(as); - goto err; } /* Init code: */ @@ -2219,7 +2320,7 @@ 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) @@ -2235,15 +2336,16 @@ void bch2_btree_root_alloc(struct bch_fs *c, enum btree_id id) closure_sync(&cl); } while (ret); - b = bch2_btree_node_mem_alloc(c); + b = bch2_btree_node_mem_alloc(c, false); bch2_btree_cache_cannibalize_unlock(c); set_btree_node_fake(b); - b->level = 0; - b->btree_id = id; + set_btree_node_need_rewrite(b); + b->c.level = 0; + b->c.btree_id = id; bkey_btree_ptr_init(&b->key); - b->key.k.p = POS_MAX; + b->key.k.p = SPOS_MAX; *((u64 *) bkey_i_to_btree_ptr(&b->key)->v.start) = U64_MAX - id; bch2_bset_init_first(b, &b->data->keys); @@ -2251,46 +2353,122 @@ void bch2_btree_root_alloc(struct bch_fs *c, enum btree_id id) b->data->flags = 0; btree_set_min(b, POS_MIN); - btree_set_max(b, POS_MAX); + btree_set_max(b, SPOS_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) +void bch2_btree_updates_to_text(struct printbuf *out, struct bch_fs *c) { - struct printbuf out = _PBUF(buf, PAGE_SIZE); struct btree_update *as; mutex_lock(&c->btree_interior_update_lock); list_for_each_entry(as, &c->btree_interior_update_list, list) - pr_buf(&out, "%p m %u w %u r %u j %llu\n", + prt_printf(out, "%p m %u w %u r %u j %llu\n", as, as->mode, as->nodes_written, atomic_read(&as->cl.remaining) & CLOSURE_REMAINING_MASK, as->journal.seq); mutex_unlock(&c->btree_interior_update_lock); - - return out.pos - buf; } -size_t bch2_btree_interior_updates_nr_pending(struct bch_fs *c) +static bool bch2_btree_interior_updates_pending(struct bch_fs *c) { - size_t ret = 0; - struct list_head *i; + bool ret; mutex_lock(&c->btree_interior_update_lock); - list_for_each(i, &c->btree_interior_update_list) - ret++; + ret = !list_empty(&c->btree_interior_update_list); mutex_unlock(&c->btree_interior_update_lock); return ret; } + +bool bch2_btree_interior_updates_flush(struct bch_fs *c) +{ + bool ret = bch2_btree_interior_updates_pending(c); + + if (ret) + closure_wait_event(&c->btree_interior_update_wait, + !bch2_btree_interior_updates_pending(c)); + return ret; +} + +void bch2_journal_entry_to_btree_root(struct bch_fs *c, struct jset_entry *entry) +{ + struct btree_root *r = &c->btree_roots[entry->btree_id]; + + mutex_lock(&c->btree_root_lock); + + 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); + + INIT_LIST_HEAD(&c->pending_node_rewrites); + mutex_init(&c->pending_node_rewrites_lock); + + 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)); +}