X-Git-Url: https://git.sesse.net/?a=blobdiff_plain;f=libbcachefs%2Fbtree_update_interior.c;h=965fdfbfa3eee032971ccdc4229847db78344423;hb=5ec29da373fa39680dcb5aa41f78b93d85fbb51f;hp=5c86e76f5079c90d79edb1253e61fa25120a3a14;hpb=f8c4bc591e8227256c54f58583de4e99432debc3;p=bcachefs-tools-debian diff --git a/libbcachefs/btree_update_interior.c b/libbcachefs/btree_update_interior.c index 5c86e76..965fdfb 100644 --- a/libbcachefs/btree_update_interior.c +++ b/libbcachefs/btree_update_interior.c @@ -16,12 +16,18 @@ #include "journal.h" #include "journal_reclaim.h" #include "keylist.h" +#include "recovery.h" #include "replicas.h" #include "super-io.h" #include #include +static void 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 *); + /* Debug code: */ /* @@ -35,11 +41,11 @@ static void btree_node_interior_verify(struct bch_fs *c, struct btree *b) struct bkey_s_c k; struct bkey_s_c_btree_ptr_v2 bp; struct bkey unpacked; - char buf1[100], buf2[100]; + struct printbuf buf1 = PRINTBUF, buf2 = PRINTBUF; BUG_ON(!b->c.level); - if (!test_bit(BCH_FS_BTREE_INTERIOR_REPLAY_DONE, &c->flags)) + if (!test_bit(JOURNAL_REPLAY_DONE, &c->journal.flags)) return; bch2_btree_node_iter_init_from_start(&iter, b); @@ -52,9 +58,9 @@ static void btree_node_interior_verify(struct bch_fs *c, struct btree *b) if (bpos_cmp(next_node, bp.v->min_key)) { bch2_dump_btree_node(c, b); - panic("expected next min_key %s got %s\n", - (bch2_bpos_to_text(&PBUF(buf1), next_node), buf1), - (bch2_bpos_to_text(&PBUF(buf2), bp.v->min_key), buf2)); + 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); @@ -62,9 +68,9 @@ static void btree_node_interior_verify(struct bch_fs *c, struct btree *b) if (bch2_btree_node_iter_end(&iter)) { if (bpos_cmp(k.k->p, b->key.k.p)) { bch2_dump_btree_node(c, b); - panic("expected end %s got %s\n", - (bch2_bpos_to_text(&PBUF(buf1), b->key.k.p), buf1), - (bch2_bpos_to_text(&PBUF(buf2), k.k->p), buf2)); + 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; } @@ -148,38 +154,26 @@ 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 *b) { - struct open_buckets ob = b->ob; + struct bch_fs *c = trans->c; + struct btree_path *path; - b->ob.nr = 0; + trans_for_each_path(trans, path) + BUG_ON(path->l[b->c.level].b == b && + path->l[b->c.level].lock_seq == b->c.lock.state.seq); - clear_btree_node_dirty(c, b); + six_lock_write(&b->c.lock, NULL, NULL); - btree_node_lock_type(c, b, SIX_LOCK_write); + bch2_btree_node_hash_remove(&c->btree_cache, b); __btree_node_free(c, b); - six_unlock_write(&b->c.lock); - - bch2_open_buckets_put(c, &ob); -} - -void bch2_btree_node_free_inmem(struct bch_fs *c, struct btree *b, - struct btree_iter *iter) -{ - struct btree_iter *linked; - trans_for_each_iter(iter->trans, linked) - BUG_ON(linked->l[b->c.level].b == b); - - six_lock_write(&b->c.lock, NULL, NULL); - __btree_node_free(c, b); six_unlock_write(&b->c.lock); six_unlock_intent(&b->c.lock); } @@ -187,6 +181,7 @@ void bch2_btree_node_free_inmem(struct bch_fs *c, struct btree *b, static struct btree *__bch2_btree_node_alloc(struct bch_fs *c, struct disk_reservation *res, struct closure *cl, + bool interior_node, unsigned flags) { struct write_point *wp; @@ -199,10 +194,10 @@ static struct btree *__bch2_btree_node_alloc(struct bch_fs *c, if (flags & BTREE_INSERT_USE_RESERVE) { nr_reserve = 0; - alloc_reserve = RESERVE_BTREE_MOVINGGC; + alloc_reserve = RESERVE_btree_movinggc; } else { nr_reserve = BTREE_NODE_RESERVE; - alloc_reserve = RESERVE_BTREE; + alloc_reserve = RESERVE_btree; } mutex_lock(&c->btree_reserve_cache_lock); @@ -230,29 +225,27 @@ retry: if (IS_ERR(wp)) return ERR_CAST(wp); - 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)); @@ -268,15 +261,19 @@ static struct btree *bch2_btree_node_alloc(struct btree_update *as, unsigned lev { 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->nr_prealloc_nodes); + BUG_ON(!p->nr); + + b = p->b[--p->nr]; - b = as->prealloc_nodes[--as->nr_prealloc_nodes]; + six_lock_intent(&b->c.lock, NULL, NULL); + six_lock_write(&b->c.lock, NULL, NULL); set_btree_node_accessed(b); - set_btree_node_dirty(c, b); + set_btree_node_dirty_acct(c, b); set_btree_node_need_write(b); bch2_bset_init_first(b, &b->data->keys); @@ -286,6 +283,7 @@ static struct btree *bch2_btree_node_alloc(struct btree_update *as, unsigned lev 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); @@ -366,7 +364,7 @@ static struct btree *__btree_root_alloc(struct btree_update *as, unsigned level) struct btree *b = bch2_btree_node_alloc(as, 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); @@ -381,70 +379,93 @@ static struct btree *__btree_root_alloc(struct btree_update *as, unsigned level) static void bch2_btree_reserve_put(struct btree_update *as) { struct bch_fs *c = as->c; + struct prealloc_nodes *p; - mutex_lock(&c->btree_reserve_cache_lock); + 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]; - while (as->nr_prealloc_nodes) { - struct btree *b = as->prealloc_nodes[--as->nr_prealloc_nodes]; + mutex_lock(&c->btree_reserve_cache_lock); - six_unlock_write(&b->c.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++]; - 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); - } + 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_type(c, b, SIX_LOCK_write); - __btree_node_free(c, b); - six_unlock_write(&b->c.lock); + mutex_unlock(&c->btree_reserve_cache_lock); - six_unlock_intent(&b->c.lock); + six_lock_intent(&b->c.lock, NULL, NULL); + six_lock_write(&b->c.lock, NULL, NULL); + __btree_node_free(c, b); + six_unlock_write(&b->c.lock); + six_unlock_intent(&b->c.lock); + } } - - mutex_unlock(&c->btree_reserve_cache_lock); } -static int bch2_btree_reserve_get(struct btree_update *as, unsigned nr_nodes, - unsigned flags, struct closure *cl) +static int bch2_btree_reserve_get(struct btree_update *as, + unsigned nr_nodes[2], + unsigned flags) { struct bch_fs *c = as->c; + struct closure cl; struct btree *b; + unsigned interior; int ret; - BUG_ON(nr_nodes > BTREE_RESERVE_MAX); + closure_init_stack(&cl); +retry: + + 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); + ret = bch2_btree_cache_cannibalize_lock(c, &cl); if (ret) - return ret; + goto err; - while (as->nr_prealloc_nodes < nr_nodes) { - b = __bch2_btree_node_alloc(c, &as->disk_res, - flags & BTREE_INSERT_NOWAIT - ? NULL : cl, flags); - if (IS_ERR(b)) { - ret = PTR_ERR(b); - goto err_free; - } + for (interior = 0; interior < 2; interior++) { + struct prealloc_nodes *p = as->prealloc_nodes + interior; + + while (p->nr < nr_nodes[interior]) { + b = __bch2_btree_node_alloc(c, &as->disk_res, + flags & BTREE_INSERT_NOWAIT + ? NULL : &cl, + interior, flags); + if (IS_ERR(b)) { + ret = PTR_ERR(b); + goto err; + } - as->prealloc_nodes[as->nr_prealloc_nodes++] = b; + p->b[p->nr++] = b; + } } bch2_btree_cache_cannibalize_unlock(c); + closure_sync(&cl); return 0; -err_free: +err: bch2_btree_cache_cannibalize_unlock(c); - trace_btree_reserve_get_fail(c, nr_nodes, cl); + closure_sync(&cl); + + if (ret == -EAGAIN) + goto retry; + + trace_btree_reserve_get_fail(c, nr_nodes[0] + nr_nodes[1], &cl); return ret; } @@ -465,31 +486,37 @@ static void bch2_btree_update_free(struct btree_update *as) bch2_disk_reservation_put(c, &as->disk_res); bch2_btree_reserve_put(as); + 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); - mutex_unlock(&c->btree_interior_update_lock); 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_will_delete_key(struct btree_update *as, - struct bkey_i *k) +static void btree_update_add_key(struct btree_update *as, + struct keylist *keys, struct btree *b) { - BUG_ON(bch2_keylist_u64s(&as->old_keys) + k->k.u64s > + struct bkey_i *k = &b->key; + + BUG_ON(bch2_keylist_u64s(keys) + k->k.u64s > ARRAY_SIZE(as->_old_keys)); - bch2_keylist_add(&as->old_keys, k); -} -static void btree_update_will_add_key(struct btree_update *as, - struct bkey_i *k) -{ - BUG_ON(bch2_keylist_u64s(&as->new_keys) + k->k.u64s > - ARRAY_SIZE(as->_new_keys)); - bch2_keylist_add(&as->new_keys, k); + bkey_copy(keys->top, k); + bkey_i_to_btree_ptr_v2(keys->top)->v.mem_ptr = b->c.level + 1; + + bch2_keylist_push(keys); } /* @@ -502,24 +529,29 @@ static int btree_update_nodes_written_trans(struct btree_trans *trans, struct bkey_i *k; int ret; - trans->extra_journal_entries = (void *) &as->journal_entries[0]; - trans->extra_journal_entry_u64s = as->journal_u64s; + ret = darray_make_room(&trans->extra_journal_entries, as->journal_u64s); + if (ret) + return ret; + + memcpy(&darray_top(trans->extra_journal_entries), + as->journal_entries, + as->journal_u64s * sizeof(u64)); + trans->extra_journal_entries.nr += as->journal_u64s; + trans->journal_pin = &as->journal; - for_each_keylist_key(&as->new_keys, k) { - ret = bch2_trans_mark_key(trans, - bkey_s_c_null, - bkey_i_to_s_c(k), - 0, 0, BTREE_TRIGGER_INSERT); + for_each_keylist_key(&as->old_keys, k) { + unsigned level = bkey_i_to_btree_ptr_v2(k)->v.mem_ptr; + + ret = bch2_trans_mark_old(trans, as->btree_id, level, bkey_i_to_s_c(k), 0); if (ret) return ret; } - for_each_keylist_key(&as->old_keys, k) { - ret = bch2_trans_mark_key(trans, - bkey_i_to_s_c(k), - bkey_s_c_null, - 0, 0, BTREE_TRIGGER_OVERWRITE); + for_each_keylist_key(&as->new_keys, k) { + unsigned level = bkey_i_to_btree_ptr_v2(k)->v.mem_ptr; + + ret = bch2_trans_mark_new(trans, as->btree_id, level, k, 0); if (ret) return ret; } @@ -547,7 +579,22 @@ static void btree_update_nodes_written(struct btree_update *as) if (ret) goto err; - BUG_ON(!journal_pin_active(&as->journal)); + /* + * 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++) { + struct btree *old = as->old_nodes[i]; + __le64 seq; + + six_lock_read(&old->c.lock, NULL, NULL); + seq = old->data ? old->data->keys.seq : 0; + six_unlock_read(&old->c.lock); + + if (seq == as->old_nodes_seq[i]) + wait_on_bit_io(&old->flags, BTREE_NODE_write_in_flight_inner, + TASK_UNINTERRUPTIBLE); + } /* * We did an update to a parent node where the pointers we added pointed @@ -567,7 +614,7 @@ static void btree_update_nodes_written(struct btree_update *as) BTREE_INSERT_NOFAIL| BTREE_INSERT_NOCHECK_RW| BTREE_INSERT_JOURNAL_RECLAIM| - BTREE_INSERT_JOURNAL_RESERVED, + JOURNAL_WATERMARK_reserved, btree_update_nodes_written_trans(&trans, as)); bch2_trans_exit(&trans); @@ -587,11 +634,13 @@ err: * we're in journal error state: */ - btree_node_lock_type(c, b, SIX_LOCK_intent); - btree_node_lock_type(c, b, SIX_LOCK_write); + six_lock_intent(&b->c.lock, NULL, NULL); + six_lock_write(&b->c.lock, NULL, NULL); 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); /* * Node might have been freed, recheck under @@ -605,8 +654,8 @@ err: if (!ret) { i->journal_seq = cpu_to_le64( - max(journal_seq, - le64_to_cpu(i->journal_seq))); + max(journal_seq, + le64_to_cpu(i->journal_seq))); bch2_btree_add_journal_pin(c, b, journal_seq); } else { @@ -636,13 +685,14 @@ err: 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); for (i = 0; i < as->nr_new_nodes; i++) { b = as->new_nodes[i]; - btree_node_lock_type(c, b, SIX_LOCK_read); + six_lock_read(&b->c.lock, NULL, NULL); btree_node_write_if_need(c, b, SIX_LOCK_read); six_unlock_read(&b->c.lock); } @@ -702,6 +752,8 @@ static void btree_update_updated_node(struct btree_update *as, struct btree *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); @@ -755,7 +807,7 @@ static void btree_update_updated_root(struct btree_update *as, struct btree *b) * And it adds @b to the list of @as's new nodes, so that we can update sector * counts in bch2_btree_update_nodes_written: */ -void bch2_btree_update_add_new_node(struct btree_update *as, struct btree *b) +static void bch2_btree_update_add_new_node(struct btree_update *as, struct btree *b) { struct bch_fs *c = as->c; @@ -767,10 +819,11 @@ void bch2_btree_update_add_new_node(struct btree_update *as, struct btree *b) as->new_nodes[as->nr_new_nodes++] = b; b->will_make_reachable = 1UL|(unsigned long) as; + set_btree_node_will_make_reachable(b); mutex_unlock(&c->btree_interior_update_lock); - btree_update_will_add_key(as, &b->key); + btree_update_add_key(as, &as->new_keys, b); } /* @@ -789,6 +842,7 @@ static void btree_update_drop_new_node(struct bch_fs *c, struct btree *b) * 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) { @@ -809,7 +863,7 @@ found: closure_put(&as->cl); } -void bch2_btree_update_get_open_buckets(struct btree_update *as, struct btree *b) +static void bch2_btree_update_get_open_buckets(struct btree_update *as, struct btree *b) { while (b->ob.nr) as->open_buckets[as->nr_open_buckets++] = @@ -821,8 +875,8 @@ void bch2_btree_update_get_open_buckets(struct btree_update *as, struct btree *b * 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 btree_update *p, *n; @@ -854,7 +908,7 @@ void bch2_btree_interior_update_will_free_node(struct btree_update *as, closure_wake_up(&c->btree_interior_update_wait); } - clear_btree_node_dirty(c, b); + clear_btree_node_dirty_acct(c, b); clear_btree_node_need_write(b); /* @@ -886,19 +940,18 @@ void bch2_btree_interior_update_will_free_node(struct btree_update *as, */ btree_update_drop_new_node(c, b); - btree_update_will_delete_key(as, &b->key); + btree_update_add_key(as, &as->old_keys, b); - /* - * XXX: Waiting on io with btree node locks held, we don't want to be - * doing this. We can't have btree writes happening after the space has - * been freed, but we really only need to block before - * btree_update_nodes_written_trans() happens. - */ - btree_node_wait_on_io(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 bch_fs *c = as->c; + u64 start_time = as->start_time; + BUG_ON(as->mode == BTREE_INTERIOR_NO_UPDATE); if (as->took_gc_lock) @@ -907,50 +960,60 @@ void bch2_btree_update_done(struct btree_update *as) bch2_btree_reserve_put(as); - continue_at(&as->cl, btree_update_set_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 btree_iter *iter, unsigned level, - unsigned nr_nodes, unsigned flags) +static struct btree_update * +bch2_btree_update_start(struct btree_trans *trans, struct btree_path *path, + unsigned level, bool split, unsigned flags) { - struct btree_trans *trans = iter->trans; struct bch_fs *c = trans->c; struct btree_update *as; - struct closure cl; + u64 start_time = local_clock(); int disk_res_flags = (flags & BTREE_INSERT_NOFAIL) ? BCH_DISK_RESERVATION_NOFAIL : 0; - int journal_flags = 0; + unsigned nr_nodes[2] = { 0, 0 }; + unsigned update_level = level; + int journal_flags = flags & JOURNAL_WATERMARK_MASK; int ret = 0; - if (flags & BTREE_INSERT_JOURNAL_RESERVED) - journal_flags |= JOURNAL_RES_GET_RESERVED; + BUG_ON(!path->should_be_locked); - closure_init_stack(&cl); -retry: - /* - * 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 (flags & BTREE_INSERT_JOURNAL_RECLAIM) + journal_flags |= JOURNAL_RES_GET_NONBLOCK; - /* - * 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); - return ERR_PTR(-EINTR); + while (1) { + nr_nodes[!!update_level] += 1 + split; + update_level++; + + if (!btree_path_node(path, update_level)) + break; + + /* + * XXX: figure out how far we might need to split, + * instead of locking/reserving all the way to the root: + */ + split = update_level + 1 < BTREE_MAX_DEPTH; + } + + /* Might have to allocate a new root: */ + if (update_level < BTREE_MAX_DEPTH) + nr_nodes[1] += 1; + + if (!bch2_btree_path_upgrade(trans, path, U8_MAX)) { + trace_trans_restart_iter_upgrade(trans->fn, _RET_IP_, + path->btree_id, &path->pos); + ret = btree_trans_restart(trans); + return ERR_PTR(ret); } if (flags & BTREE_INSERT_GC_LOCK_HELD) lockdep_assert_held(&c->gc_lock); else if (!down_read_trylock(&c->gc_lock)) { - if (flags & BTREE_INSERT_NOUNLOCK) - return ERR_PTR(-EINTR); - bch2_trans_unlock(trans); down_read(&c->gc_lock); if (!bch2_trans_relock(trans)) { @@ -963,9 +1026,10 @@ retry: 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->took_gc_lock = !(flags & BTREE_INSERT_GC_LOCK_HELD); - as->btree_id = iter->btree_id; + as->btree_id = path->btree_id; INIT_LIST_HEAD(&as->list); INIT_LIST_HEAD(&as->unwritten_list); INIT_LIST_HEAD(&as->write_blocked_list); @@ -973,76 +1037,53 @@ retry: bch2_keylist_init(&as->new_keys, as->_new_keys); bch2_keylist_init(&as->parent_keys, as->inline_keys); - ret = bch2_journal_preres_get(&c->journal, &as->journal_preres, - BTREE_UPDATE_JOURNAL_RES, - journal_flags|JOURNAL_RES_GET_NONBLOCK); - if (ret == -EAGAIN) { - /* - * this would be cleaner if bch2_journal_preres_get() took a - * closure argument - */ - if (flags & BTREE_INSERT_NOUNLOCK) { - trace_trans_restart_journal_preres_get(trans->ip); - ret = -EINTR; - goto err; - } - - bch2_trans_unlock(trans); + mutex_lock(&c->btree_interior_update_lock); + list_add_tail(&as->list, &c->btree_interior_update_list); + mutex_unlock(&c->btree_interior_update_lock); - if (flags & BTREE_INSERT_JOURNAL_RECLAIM) { - bch2_btree_update_free(as); - return ERR_PTR(ret); - } + /* + * 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); - if (ret) { - trace_trans_restart_journal_preres_get(trans->ip); - goto err; - } + bch2_trans_unlock(trans); - if (!bch2_trans_relock(trans)) { - ret = -EINTR; - goto err; - } + ret = bch2_journal_preres_get(&c->journal, &as->journal_preres, + BTREE_UPDATE_JOURNAL_RES, + journal_flags); + if (ret) { + bch2_btree_update_free(as); + trace_trans_restart_journal_preres_get(trans->fn, _RET_IP_); + btree_trans_restart(trans); + return ERR_PTR(ret); } ret = bch2_disk_reservation_get(c, &as->disk_res, - nr_nodes * c->opts.btree_node_size, + (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(as, nr_nodes, flags, - !(flags & BTREE_INSERT_NOUNLOCK) ? &cl : NULL); + ret = bch2_btree_reserve_get(as, nr_nodes, flags); if (ret) goto err; - bch2_journal_pin_add(&c->journal, - atomic64_read(&c->journal.seq), - &as->journal, NULL); - - mutex_lock(&c->btree_interior_update_lock); - list_add_tail(&as->list, &c->btree_interior_update_list); - mutex_unlock(&c->btree_interior_update_lock); + if (!bch2_trans_relock(trans)) { + ret = -EINTR; + goto err; + } return as; err: bch2_btree_update_free(as); - - if (ret == -EAGAIN) { - BUG_ON(flags & BTREE_INSERT_NOUNLOCK); - - bch2_trans_unlock(trans); - closure_sync(&cl); - ret = -EINTR; - } - - if (ret == -EINTR && bch2_trans_relock(trans)) - goto retry; - return ERR_PTR(ret); } @@ -1055,11 +1096,6 @@ static void bch2_btree_set_root_inmem(struct bch_fs *c, struct btree *b) list_del_init(&b->list); mutex_unlock(&c->btree_cache.lock); - if (b->c.level) - six_lock_pcpu_alloc(&b->c.lock); - else - six_lock_pcpu_free(&b->c.lock); - mutex_lock(&c->btree_root_lock); BUG_ON(btree_node_root(c, b) && (b->c.level < btree_node_root(c, b)->c.level || @@ -1083,15 +1119,16 @@ static void bch2_btree_set_root_inmem(struct bch_fs *c, struct btree *b) * 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)); + BUG_ON(!b->written); old = btree_node_root(c, b); @@ -1099,7 +1136,7 @@ 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(trans, path, old); bch2_btree_set_root_inmem(c, b); @@ -1112,27 +1149,40 @@ 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 bkey_packed *k; - const char *invalid; - - invalid = bch2_bkey_invalid(c, bkey_i_to_s_c(insert), btree_node_type(b)) ?: - bch2_bkey_in_btree_node(b, bkey_i_to_s_c(insert)); - if (invalid) { - char buf[160]; - - bch2_bkey_val_to_text(&PBUF(buf), c, bkey_i_to_s_c(insert)); - bch2_fs_inconsistent(c, "inserting invalid bkey %s: %s", buf, invalid); + struct printbuf buf = PRINTBUF; + + 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(); } @@ -1149,9 +1199,35 @@ static void bch2_insert_fixup_btree_ptr(struct btree_update *as, struct btree *b bkey_iter_pos_cmp(b, k, &insert->k.p) < 0) bch2_btree_node_iter_advance(node_iter, b); - bch2_btree_bset_insert_key(iter, b, node_iter, insert); - set_btree_node_dirty(c, b); + bch2_btree_bset_insert_key(trans, path, b, node_iter, insert); + set_btree_node_dirty_acct(c, 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)) { + bch2_insert_fixup_btree_ptr(as, trans, path, b, + &node_iter, bch2_keylist_front(keys)); + bch2_keylist_pop_front(keys); + } } /* @@ -1159,8 +1235,7 @@ static void bch2_insert_fixup_btree_ptr(struct btree_update *as, struct btree *b * node) */ static struct btree *__btree_split_node(struct btree_update *as, - struct btree *n1, - struct btree_iter *iter) + struct btree *n1) { struct bkey_format_state s; size_t nr_packed = 0, nr_unpacked = 0; @@ -1170,13 +1245,14 @@ static struct btree *__btree_split_node(struct btree_update *as, struct bpos n1_pos; n2 = bch2_btree_node_alloc(as, n1->c.level); - bch2_btree_update_add_new_node(as, n2); n2->data->max_key = n1->data->max_key; n2->data->format = n1->format; SET_BTREE_NODE_SEQ(n2->data, BTREE_NODE_SEQ(n1->data)); n2->key.k.p = n1->key.k.p; + bch2_btree_update_add_new_node(as, n2); + set1 = btree_bset_first(n1); set2 = btree_bset_first(n2); @@ -1275,8 +1351,10 @@ 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; @@ -1284,16 +1362,9 @@ static void btree_split_insert_keys(struct btree_update *as, struct btree *b, 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); - - 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 @@ -1323,37 +1394,39 @@ static void btree_split_insert_keys(struct btree_update *as, struct btree *b, btree_node_interior_verify(as->c, b); } -static void btree_split(struct btree_update *as, struct btree *b, - struct btree_iter *iter, struct keylist *keys, - unsigned flags) +static void 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; u64 start_time = local_clock(); BUG_ON(!parent && (b != btree_node_root(c, b))); - BUG_ON(!btree_node_intent_locked(iter, btree_node_root(c, b)->c.level)); + BUG_ON(!btree_node_intent_locked(path, btree_node_root(c, b)->c.level)); bch2_btree_interior_update_will_free_node(as, b); n1 = bch2_btree_node_alloc_replacement(as, b); - bch2_btree_update_add_new_node(as, n1); if (keys) - btree_split_insert_keys(as, n1, iter, keys); + btree_split_insert_keys(as, trans, path, n1, keys); if (bset_u64s(&n1->set[0]) > BTREE_SPLIT_THRESHOLD(c)) { trace_btree_split(c, b); - n2 = __btree_split_node(as, n1, iter); + n2 = __btree_split_node(as, n1); bch2_btree_build_aux_trees(n2); bch2_btree_build_aux_trees(n1); six_unlock_write(&n2->c.lock); six_unlock_write(&n1->c.lock); - bch2_btree_node_write(c, n2, SIX_LOCK_intent); + bch2_btree_update_add_new_node(as, n1); + + bch2_btree_node_write(c, n1, SIX_LOCK_intent, 0); + bch2_btree_node_write(c, n2, SIX_LOCK_intent, 0); /* * Note that on recursive parent_keys == keys, so we @@ -1370,9 +1443,9 @@ static void btree_split(struct btree_update *as, struct btree *b, n3->sib_u64s[0] = U16_MAX; n3->sib_u64s[1] = U16_MAX; - btree_split_insert_keys(as, n3, iter, &as->parent_keys); + btree_split_insert_keys(as, trans, path, n3, &as->parent_keys); - bch2_btree_node_write(c, n3, SIX_LOCK_intent); + bch2_btree_node_write(c, n3, SIX_LOCK_intent, 0); } } else { trace_btree_compact(c, b); @@ -1380,22 +1453,24 @@ static void btree_split(struct btree_update *as, struct btree *b, bch2_btree_build_aux_trees(n1); six_unlock_write(&n1->c.lock); + bch2_btree_update_add_new_node(as, n1); + + bch2_btree_node_write(c, n1, SIX_LOCK_intent, 0); + if (parent) bch2_keylist_add(&as->parent_keys, &n1->key); } - bch2_btree_node_write(c, n1, SIX_LOCK_intent); - /* 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); + bch2_btree_insert_node(as, trans, path, parent, &as->parent_keys, flags); } 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_btree_update_get_open_buckets(as, n1); @@ -1404,15 +1479,14 @@ static void btree_split(struct btree_update *as, struct btree *b, if (n3) bch2_btree_update_get_open_buckets(as, n3); - /* Successful split, update the iterator to point to the new nodes: */ + /* Successful split, update the path to point to the new nodes: */ six_lock_increment(&b->c.lock, SIX_LOCK_intent); - bch2_btree_iter_node_drop(iter, b); if (n3) - bch2_btree_iter_node_replace(iter, n3); + bch2_trans_node_add(trans, n3); if (n2) - bch2_btree_iter_node_replace(iter, n2); - bch2_btree_iter_node_replace(iter, n1); + bch2_trans_node_add(trans, n2); + bch2_trans_node_add(trans, n1); /* * The old node must be freed (in memory) _before_ unlocking the new @@ -1420,7 +1494,7 @@ 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, b); if (n3) six_unlock_intent(&n3->c.lock); @@ -1428,42 +1502,32 @@ static void btree_split(struct btree_update *as, struct btree *b, six_unlock_intent(&n2->c.lock); six_unlock_intent(&n1->c.lock); - 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); } 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->c.level].iter; - - /* - * 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_left_packed(b, k, &insert->k.p) >= 0)) - ; - - for_each_keylist_key(keys, insert) - bch2_insert_fixup_btree_ptr(as, b, iter, insert, &node_iter); + __bch2_btree_insert_keys_interior(as, trans, path, b, + path->l[b->c.level].iter, keys); btree_update_updated_node(as, b); - trans_for_each_iter_with_node(iter->trans, b, linked) + trans_for_each_path_with_node(trans, b, linked) bch2_btree_node_iter_peek(&linked->l[b->c.level].iter, b); - bch2_btree_trans_verify_iters(iter->trans, b); + bch2_trans_verify_paths(trans); } /** @@ -1478,9 +1542,9 @@ 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 void 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); @@ -1488,21 +1552,21 @@ void bch2_btree_insert_node(struct btree_update *as, struct btree *b, int live_u64s_added, u64s_added; lockdep_assert_held(&c->gc_lock); - BUG_ON(!btree_node_intent_locked(iter, btree_node_root(c, b)->c.level)); + BUG_ON(!btree_node_intent_locked(path, btree_node_root(c, b)->c.level)); BUG_ON(!b->c.level); BUG_ON(!as || as->b); bch2_verify_keylist_sorted(keys); - bch2_btree_node_lock_for_insert(c, b, iter); + bch2_btree_node_lock_for_insert(trans, path, b); if (!bch2_btree_node_insert_fits(c, b, bch2_keylist_u64s(keys))) { - bch2_btree_node_unlock_write(b, iter); + bch2_btree_node_unlock_write(trans, path, b); goto split; } btree_node_interior_verify(c, b); - bch2_btree_insert_keys_interior(as, b, iter, keys); + 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; @@ -1514,46 +1578,47 @@ 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_trans_node_reinit_iter(trans, b); - bch2_btree_node_unlock_write(b, iter); + bch2_btree_node_unlock_write(trans, path, b); btree_node_interior_verify(c, b); return; split: - btree_split(as, b, iter, keys, flags); + 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 *b = iter_l(iter)->b; + struct btree *b = path_l(path)->b; struct btree_update *as; unsigned l; int ret = 0; - as = bch2_btree_update_start(iter, iter->level, - btree_update_reserve_required(c, b), flags); + as = bch2_btree_update_start(trans, path, path->level, + true, flags); if (IS_ERR(as)) return PTR_ERR(as); - btree_split(as, b, iter, NULL, flags); + btree_split(as, trans, path, b, NULL, flags); bch2_btree_update_done(as); - for (l = iter->level + 1; btree_iter_node(iter, l) && !ret; l++) - ret = bch2_foreground_maybe_merge(c, iter, l, flags); + for (l = path->level + 1; btree_path_node(path, l) && !ret; l++) + ret = bch2_foreground_maybe_merge(trans, path, l, flags); return ret; } -int __bch2_foreground_maybe_merge(struct bch_fs *c, - struct btree_iter *iter, +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 btree_iter *sib_iter = NULL; + struct bch_fs *c = trans->c; + struct btree_path *sib_path = NULL; struct btree_update *as; struct bkey_format_state new_s; struct bkey_format new_f; @@ -1561,39 +1626,36 @@ int __bch2_foreground_maybe_merge(struct bch_fs *c, struct btree *b, *m, *n, *prev, *next, *parent; struct bpos sib_pos; size_t sib_u64s; - int ret = 0, ret2 = 0; - - BUG_ON(!btree_node_locked(iter, level)); -retry: - ret = bch2_btree_iter_traverse(iter); - if (ret) - goto err; + u64 start_time = local_clock(); + int ret = 0; - 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; if ((sib == btree_prev_sib && !bpos_cmp(b->data->min_key, POS_MIN)) || - (sib == btree_next_sib && !bpos_cmp(b->data->max_key, POS_MAX))) { + (sib == btree_next_sib && !bpos_cmp(b->data->max_key, SPOS_MAX))) { b->sib_u64s[sib] = U16_MAX; - goto out; + return 0; } sib_pos = sib == btree_prev_sib ? bpos_predecessor(b->data->min_key) : bpos_successor(b->data->max_key); - sib_iter = bch2_trans_get_node_iter(trans, iter->btree_id, - sib_pos, U8_MAX, level, - BTREE_ITER_INTENT); - ret = bch2_btree_iter_traverse(sib_iter); + 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; - m = sib_iter->l[level].b; + sib_path->should_be_locked = true; - if (btree_node_parent(iter, b) != - btree_node_parent(sib_iter, m)) { + 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; } @@ -1607,15 +1669,18 @@ retry: } if (bkey_cmp(bpos_successor(prev->data->max_key), next->data->min_key)) { - char buf1[100], buf2[100]; - - bch2_bpos_to_text(&PBUF(buf1), prev->data->max_key); - bch2_bpos_to_text(&PBUF(buf2), next->data->min_key); - bch2_fs_inconsistent(c, - "btree topology error in btree merge:\n" - "prev ends at %s\n" - "next starts at %s\n", - buf1, buf2); + 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, + "btree topology error in btree merge:\n" + " prev ends at %s\n" + " next starts at %s", + buf1.buf, buf2.buf); + printbuf_exit(&buf1); + printbuf_exit(&buf2); + bch2_topology_error(c); ret = -EIO; goto err; } @@ -1643,12 +1708,11 @@ retry: if (b->sib_u64s[sib] > c->btree_foreground_merge_threshold) goto out; - parent = btree_node_parent(iter, b); - as = bch2_btree_update_start(iter, level, - btree_update_reserve_required(c, parent) + 1, - flags| + parent = btree_node_parent(path, b); + as = bch2_btree_update_start(trans, path, level, false, BTREE_INSERT_NOFAIL| - BTREE_INSERT_USE_RESERVE); + BTREE_INSERT_USE_RESERVE| + flags); ret = PTR_ERR_OR_ZERO(as); if (ret) goto err; @@ -1659,12 +1723,17 @@ retry: bch2_btree_interior_update_will_free_node(as, m); n = bch2_btree_node_alloc(as, b->c.level); - bch2_btree_update_add_new_node(as, n); + + 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; + bch2_btree_update_add_new_node(as, n); + + n->data->format = new_f; btree_node_set_format(n, new_f); bch2_btree_sort_into(c, n, prev); @@ -1673,96 +1742,64 @@ retry: bch2_btree_build_aux_trees(n); six_unlock_write(&n->c.lock); + bch2_btree_node_write(c, n, SIX_LOCK_intent, 0); + 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, trans, path, parent, &as->parent_keys, flags); - bch2_btree_insert_node(as, parent, iter, &as->parent_keys, flags); + bch2_trans_verify_paths(trans); bch2_btree_update_get_open_buckets(as, n); six_lock_increment(&b->c.lock, SIX_LOCK_intent); six_lock_increment(&m->c.lock, SIX_LOCK_intent); - bch2_btree_iter_node_drop(iter, b); - bch2_btree_iter_node_drop(iter, m); - bch2_btree_iter_node_replace(iter, n); + bch2_trans_node_add(trans, n); - bch2_btree_trans_verify_iters(trans, n); + bch2_trans_verify_paths(trans); - bch2_btree_node_free_inmem(c, b, iter); - bch2_btree_node_free_inmem(c, m, iter); + bch2_btree_node_free_inmem(trans, b); + bch2_btree_node_free_inmem(trans, m); six_unlock_intent(&n->c.lock); bch2_btree_update_done(as); -out: - bch2_btree_trans_verify_locks(trans); - bch2_trans_iter_free(trans, sib_iter); - /* - * 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: - */ - return ret ?: ret2; + bch2_time_stats_update(&c->times[BCH_TIME_btree_node_merge], start_time); +out: err: - bch2_trans_iter_put(trans, sib_iter); - sib_iter = NULL; - - if (ret == -EINTR && bch2_trans_relock(trans)) - goto retry; - - if (ret == -EINTR && !(flags & BTREE_INSERT_NOUNLOCK)) { - ret2 = ret; - ret = bch2_btree_iter_traverse_all(trans); - if (!ret) - goto retry; - } - - goto out; + bch2_path_put(trans, sib_path, true); + bch2_trans_verify_locks(trans); + return ret; } /** * bch_btree_node_rewrite - Rewrite/move a btree node */ -int bch2_btree_node_rewrite(struct bch_fs *c, struct btree_iter *iter, - __le64 seq, unsigned flags) +int bch2_btree_node_rewrite(struct btree_trans *trans, + struct btree_iter *iter, + struct btree *b, + unsigned flags) { - struct btree *b, *n, *parent; + struct bch_fs *c = trans->c; + struct btree *n, *parent; struct btree_update *as; int ret; flags |= BTREE_INSERT_NOFAIL; -retry: - ret = bch2_btree_iter_traverse(iter); - if (ret) - goto out; - - b = bch2_btree_iter_peek_node(iter); - if (!b || b->data->keys.seq != seq) - goto out; - parent = btree_node_parent(iter, b); - as = bch2_btree_update_start(iter, b->c.level, - (parent - ? btree_update_reserve_required(c, parent) - : 0) + 1, - flags); + 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 == -EINTR) - goto retry; - if (ret) { - trace_btree_gc_rewrite_node_fail(c, b); + if (ret) goto out; - } bch2_btree_interior_update_will_free_node(as, b); @@ -1772,100 +1809,219 @@ retry: bch2_btree_build_aux_trees(n); six_unlock_write(&n->c.lock); - trace_btree_gc_rewrite_node(c, b); + trace_btree_rewrite(c, b); - bch2_btree_node_write(c, n, SIX_LOCK_intent); + bch2_btree_node_write(c, n, SIX_LOCK_intent, 0); if (parent) { bch2_keylist_add(&as->parent_keys, &n->key); - bch2_btree_insert_node(as, parent, iter, &as->parent_keys, flags); + bch2_btree_insert_node(as, trans, iter->path, parent, + &as->parent_keys, flags); } else { - bch2_btree_set_root(as, n, iter); + bch2_btree_set_root(as, trans, iter->path, n); } bch2_btree_update_get_open_buckets(as, n); six_lock_increment(&b->c.lock, SIX_LOCK_intent); - bch2_btree_iter_node_drop(iter, b); - bch2_btree_iter_node_replace(iter, n); - bch2_btree_node_free_inmem(c, b, iter); + bch2_trans_node_add(trans, n); + bch2_btree_node_free_inmem(trans, b); six_unlock_intent(&n->c.lock); bch2_btree_update_done(as); out: - bch2_btree_iter_downgrade(iter); + bch2_btree_path_downgrade(iter->path); + return ret; +} + +struct async_btree_rewrite { + struct bch_fs *c; + struct work_struct work; + 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_iter iter; + struct btree *b; + int ret; + + 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; + + if (!b || b->data->keys.seq != a->seq) + goto out; + + ret = bch2_btree_node_rewrite(trans, &iter, b, 0); +out : + bch2_trans_iter_exit(trans, &iter); + return ret; } -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) +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; + + bch2_trans_do(c, NULL, NULL, 0, + async_btree_node_rewrite_trans(&trans, a)); + percpu_ref_put(&c->writes); + kfree(a); +} + +void bch2_btree_node_rewrite_async(struct bch_fs *c, struct btree *b) +{ + struct async_btree_rewrite *a; + + if (!percpu_ref_tryget_live(&c->writes)) + return; + + a = kmalloc(sizeof(*a), GFP_NOFS); + if (!a) { + percpu_ref_put(&c->writes); + 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); + queue_work(c->btree_interior_update_worker, &a->work); +} + +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; - btree_update_will_delete_key(as, &b->key); - btree_update_will_add_key(as, new_key); + 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; + + ret = bch2_trans_mark_new(trans, b->c.btree_id, b->c.level + 1, + new_key, 0); + if (ret) + return 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->c.level, b->c.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_cmp(iter2.path->pos, new_key->k.p)); - bch2_btree_node_hash_remove(&c->btree_cache, b); + btree_node_unlock(iter2.path, iter2.path->level); + path_l(iter2.path)->b = BTREE_ITER_NO_NODE_UP; + iter2.path->level++; + btree_path_set_dirty(iter2.path, BTREE_ITER_NEED_TRAVERSE); - 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_path_check_sort(trans, iter2.path, 0); + + 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); - bkey_copy(&b->key, new_key); + ret = darray_make_room(&trans->extra_journal_entries, + jset_u64s(new_key->k.u64s)); + if (ret) + return ret; - if (btree_ptr_hash_val(&b->key) != b->hash_val) { - mutex_lock(&c->btree_cache.lock); - bch2_btree_node_hash_remove(&c->btree_cache, b); + 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); + } - ret = __bch2_btree_node_hash_insert(&c->btree_cache, b); - BUG_ON(ret); - mutex_unlock(&c->btree_cache.lock); - } + 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; + + bch2_btree_node_lock_write(trans, iter->path, b); + + 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); - btree_update_updated_root(as, b); - bch2_btree_node_unlock_write(b, iter); + 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 = 0; + if (!btree_node_intent_locked(path, b->c.level) && + !bch2_btree_path_upgrade(trans, path, b->c.level + 1)) { + btree_trans_restart(trans); + return -EINTR; + } + closure_init_stack(&cl); /* @@ -1875,27 +2031,20 @@ int bch2_btree_node_update_key(struct bch_fs *c, struct btree_iter *iter, if (btree_ptr_hash_val(new_key) != b->hash_val) { ret = bch2_btree_cache_cannibalize_lock(c, &cl); if (ret) { - bch2_trans_unlock(iter->trans); + bch2_trans_unlock(trans); closure_sync(&cl); - if (!bch2_trans_relock(iter->trans)) + if (!bch2_trans_relock(trans)) return -EINTR; } - new_hash = bch2_btree_node_mem_alloc(c); + new_hash = bch2_btree_node_mem_alloc(c, false); } - as = bch2_btree_update_start(iter, b->c.level, - parent ? btree_update_reserve_required(c, parent) : 0, - BTREE_INSERT_NOFAIL); - if (IS_ERR(as)) { - ret = PTR_ERR(as); - goto err; - } + path->intent_ref++; + ret = __bch2_btree_node_update_key(trans, iter, b, new_hash, + new_key, skip_triggers); + --path->intent_ref; - __bch2_btree_node_update_key(c, as, iter, b, new_hash, new_key); - - bch2_btree_iter_downgrade(iter); -err: if (new_hash) { mutex_lock(&c->btree_cache.lock); list_move(&new_hash->list, &c->btree_cache.freeable); @@ -1909,6 +2058,35 @@ err: 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; +} + /* Init code: */ /* @@ -1935,7 +2113,7 @@ 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); @@ -1944,7 +2122,7 @@ void bch2_btree_root_alloc(struct bch_fs *c, enum btree_id id) 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); @@ -1952,7 +2130,7 @@ 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); @@ -1972,7 +2150,7 @@ void bch2_btree_updates_to_text(struct printbuf *out, struct bch_fs *c) 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, @@ -1981,19 +2159,27 @@ void bch2_btree_updates_to_text(struct printbuf *out, struct bch_fs *c) mutex_unlock(&c->btree_interior_update_lock); } -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_entries_to_btree_roots(struct bch_fs *c, struct jset *jset) { struct btree_root *r;