X-Git-Url: https://git.sesse.net/?a=blobdiff_plain;f=libbcachefs%2Fbtree_update_leaf.c;h=f5160d4f8192eb19e649631c70d6a691180f7126;hb=b422ff58ba8eedcfef3b67b66468660f07b0cfc1;hp=6e9688d0bb77324d10e6834fb4c42a0aebc8738e;hpb=3b3574e8f64b2485a72bb2ab18e7a6ca0ec3de8e;p=bcachefs-tools-debian diff --git a/libbcachefs/btree_update_leaf.c b/libbcachefs/btree_update_leaf.c index 6e9688d..f5160d4 100644 --- a/libbcachefs/btree_update_leaf.c +++ b/libbcachefs/btree_update_leaf.c @@ -21,6 +21,14 @@ #include #include +static inline int btree_insert_entry_cmp(const struct btree_insert_entry *l, + const struct btree_insert_entry *r) +{ + return cmp_int(l->btree_id, r->btree_id) ?: + -cmp_int(l->level, r->level) ?: + bpos_cmp(l->k->k.p, r->k->k.p); +} + static inline bool same_leaf_as_prev(struct btree_trans *trans, struct btree_insert_entry *i) { @@ -62,27 +70,24 @@ bool bch2_btree_bset_insert_key(struct btree_iter *iter, EBUG_ON(btree_node_just_written(b)); EBUG_ON(bset_written(b, btree_bset_last(b))); EBUG_ON(bkey_deleted(&insert->k) && bkey_val_u64s(&insert->k)); - EBUG_ON(bkey_cmp(b->data->min_key, POS_MIN) && - bkey_cmp(bkey_start_pos(&insert->k), - bkey_predecessor(b->data->min_key)) < 0); - EBUG_ON(bkey_cmp(insert->k.p, b->data->min_key) < 0); - EBUG_ON(bkey_cmp(insert->k.p, b->data->max_key) > 0); + EBUG_ON(bpos_cmp(insert->k.p, b->data->min_key) < 0); + EBUG_ON(bpos_cmp(insert->k.p, b->data->max_key) > 0); EBUG_ON(insert->k.u64s > bch_btree_keys_u64s_remaining(iter->trans->c, b)); EBUG_ON(iter->flags & BTREE_ITER_IS_EXTENTS); k = bch2_btree_node_iter_peek_all(node_iter, b); - if (k && bkey_cmp_packed(b, k, &insert->k)) + if (k && bkey_cmp_left_packed(b, k, &insert->k.p)) k = NULL; /* @k is the key being overwritten/deleted, if any: */ - EBUG_ON(k && bkey_whiteout(k)); + EBUG_ON(k && bkey_deleted(k)); /* Deleting, but not found? nothing to do: */ - if (bkey_whiteout(&insert->k) && !k) + if (bkey_deleted(&insert->k) && !k) return false; - if (bkey_whiteout(&insert->k)) { + if (bkey_deleted(&insert->k)) { /* Deleting: */ btree_account_key_drop(b, k); k->type = KEY_TYPE_deleted; @@ -129,7 +134,7 @@ fix_iter: return true; } -static void __btree_node_flush(struct journal *j, struct journal_entry_pin *pin, +static int __btree_node_flush(struct journal *j, struct journal_entry_pin *pin, unsigned i, u64 seq) { struct bch_fs *c = container_of(j, struct bch_fs, journal); @@ -140,14 +145,15 @@ static void __btree_node_flush(struct journal *j, struct journal_entry_pin *pin, bch2_btree_node_write_cond(c, b, (btree_current_write(b) == w && w->journal.seq == seq)); six_unlock_read(&b->c.lock); + return 0; } -static void btree_node_flush0(struct journal *j, struct journal_entry_pin *pin, u64 seq) +static int btree_node_flush0(struct journal *j, struct journal_entry_pin *pin, u64 seq) { return __btree_node_flush(j, pin, 0, seq); } -static void btree_node_flush1(struct journal *j, struct journal_entry_pin *pin, u64 seq) +static int btree_node_flush1(struct journal *j, struct journal_entry_pin *pin, u64 seq) { return __btree_node_flush(j, pin, 1, seq); } @@ -191,7 +197,7 @@ static bool btree_insert_key_leaf(struct btree_trans *trans, bch2_btree_add_journal_pin(c, b, trans->journal_res.seq); if (unlikely(!btree_node_dirty(b))) - set_btree_node_dirty(b); + set_btree_node_dirty(c, b); live_u64s_added = (int) b->nr.live_u64s - old_live_u64s; u64s_added = (int) bset_u64s(t) - old_u64s; @@ -214,15 +220,23 @@ static bool btree_insert_key_leaf(struct btree_trans *trans, /* Normal update interface: */ static inline void btree_insert_entry_checks(struct btree_trans *trans, - struct btree_iter *iter, - struct bkey_i *insert) + struct btree_insert_entry *i) { struct bch_fs *c = trans->c; - BUG_ON(bkey_cmp(insert->k.p, iter->pos)); - BUG_ON(debug_check_bkeys(c) && - bch2_bkey_invalid(c, bkey_i_to_s_c(insert), - __btree_node_type(iter->level, iter->btree_id))); + if (bch2_debug_check_bkeys) { + const char *invalid = bch2_bkey_invalid(c, + bkey_i_to_s_c(i->k), i->bkey_type); + if (invalid) { + char buf[200]; + + bch2_bkey_val_to_text(&PBUF(buf), c, bkey_i_to_s_c(i->k)); + panic("invalid bkey %s on insert: %s\n", buf, invalid); + } + } + BUG_ON(!i->is_extent && bpos_cmp(i->k->k.p, i->iter->real_pos)); + BUG_ON(i->level != i->iter->level); + BUG_ON(i->btree_id != i->iter->btree_id); } static noinline int @@ -264,30 +278,12 @@ static inline int bch2_trans_journal_res_get(struct btree_trans *trans, static enum btree_insert_ret btree_key_can_insert(struct btree_trans *trans, struct btree_iter *iter, - struct bkey_i *insert, - unsigned *u64s) + unsigned u64s) { struct bch_fs *c = trans->c; struct btree *b = iter_l(iter)->b; - static enum btree_insert_ret ret; - - if (unlikely(btree_node_fake(b))) - return BTREE_INSERT_BTREE_NODE_FULL; - - /* - * old bch2_extent_sort_fix_overlapping() algorithm won't work with new - * style extent updates: - */ - if (unlikely(btree_node_old_extent_overwrite(b))) - return BTREE_INSERT_BTREE_NODE_FULL; - - ret = !btree_iter_is_extents(iter) - ? BTREE_INSERT_OK - : bch2_extent_can_insert(trans, iter, insert); - if (ret) - return ret; - if (*u64s > bch_btree_keys_u64s_remaining(c, b)) + if (!bch2_btree_node_insert_fits(c, b, u64s)) return BTREE_INSERT_BTREE_NODE_FULL; return BTREE_INSERT_OK; @@ -296,8 +292,7 @@ btree_key_can_insert(struct btree_trans *trans, static enum btree_insert_ret btree_key_can_insert_cached(struct btree_trans *trans, struct btree_iter *iter, - struct bkey_i *insert, - unsigned *u64s) + unsigned u64s) { struct bkey_cached *ck = (void *) iter->l[0].b; unsigned new_u64s; @@ -305,10 +300,15 @@ btree_key_can_insert_cached(struct btree_trans *trans, BUG_ON(iter->level); - if (*u64s <= ck->u64s) + if (!test_bit(BKEY_CACHED_DIRTY, &ck->flags) && + bch2_btree_key_cache_must_wait(trans->c) && + !(trans->flags & BTREE_INSERT_JOURNAL_RECLAIM)) + return BTREE_INSERT_NEED_JOURNAL_RECLAIM; + + if (u64s <= ck->u64s) return BTREE_INSERT_OK; - new_u64s = roundup_pow_of_two(*u64s); + new_u64s = roundup_pow_of_two(u64s); new_k = krealloc(ck->k, new_u64s * sizeof(u64), GFP_NOFS); if (!new_k) return -ENOMEM; @@ -349,18 +349,6 @@ static inline void do_btree_insert_one(struct btree_trans *trans, } } -static inline bool iter_has_trans_triggers(struct btree_iter *iter) -{ - return BTREE_NODE_TYPE_HAS_TRANS_TRIGGERS & (1U << iter->btree_id); -} - -static inline bool iter_has_nontrans_triggers(struct btree_iter *iter) -{ - return (BTREE_NODE_TYPE_HAS_TRIGGERS & - ~BTREE_NODE_TYPE_HAS_TRANS_TRIGGERS) & - (1U << iter->btree_id); -} - static noinline void bch2_btree_iter_unlock_noinline(struct btree_iter *iter) { __bch2_btree_iter_unlock(iter); @@ -388,8 +376,8 @@ bch2_trans_commit_write_locked(struct btree_trans *trans, struct btree_insert_entry **stopped_at) { struct bch_fs *c = trans->c; - struct bch_fs_usage *fs_usage = NULL; struct btree_insert_entry *i; + struct btree_trans_commit_hook *h; unsigned u64s = 0; bool marking = false; int ret; @@ -407,6 +395,14 @@ bch2_trans_commit_write_locked(struct btree_trans *trans, prefetch(&trans->c->journal.flags); + h = trans->hooks; + while (h) { + ret = h->fn(trans, h); + if (ret) + return ret; + h = h->next; + } + trans_for_each_update2(trans, i) { /* Multiple inserts might go to same leaf: */ if (!same_leaf_as_prev(trans, i)) @@ -414,20 +410,26 @@ bch2_trans_commit_write_locked(struct btree_trans *trans, u64s += i->k->k.u64s; ret = btree_iter_type(i->iter) != BTREE_ITER_CACHED - ? btree_key_can_insert(trans, i->iter, i->k, &u64s) - : btree_key_can_insert_cached(trans, i->iter, i->k, &u64s); + ? btree_key_can_insert(trans, i->iter, u64s) + : btree_key_can_insert_cached(trans, i->iter, u64s); if (ret) { *stopped_at = i; return ret; } - if (btree_node_type_needs_gc(i->iter->btree_id)) + if (btree_node_type_needs_gc(i->bkey_type)) marking = true; } if (marking) { percpu_down_read(&c->mark_lock); - fs_usage = bch2_fs_usage_scratch_get(c); + } + + /* Must be called under mark_lock: */ + if (marking && trans->fs_usage_deltas && + !bch2_replicas_delta_list_marked(c, trans->fs_usage_deltas)) { + ret = BTREE_INSERT_NEED_MARK_REPLICAS; + goto err; } /* @@ -458,29 +460,21 @@ bch2_trans_commit_write_locked(struct btree_trans *trans, */ if (!(trans->flags & BTREE_INSERT_JOURNAL_REPLAY)) { - if (journal_seq_verify(c)) + if (bch2_journal_seq_verify) trans_for_each_update2(trans, i) i->k->k.version.lo = trans->journal_res.seq; - else if (inject_invalid_keys(c)) + else if (bch2_inject_invalid_keys) trans_for_each_update2(trans, i) i->k->k.version = MAX_VERSION; } - /* Must be called under mark_lock: */ - if (marking && trans->fs_usage_deltas && - bch2_replicas_delta_list_apply(c, fs_usage, - trans->fs_usage_deltas)) { - ret = BTREE_INSERT_NEED_MARK_REPLICAS; - goto err; - } - trans_for_each_update(trans, i) - if (iter_has_nontrans_triggers(i->iter)) + if (BTREE_NODE_TYPE_HAS_MEM_TRIGGERS & (1U << i->bkey_type)) bch2_mark_update(trans, i->iter, i->k, - fs_usage, i->trigger_flags); + NULL, i->trigger_flags); - if (marking) - bch2_trans_fs_usage_apply(trans, fs_usage); + if (marking && trans->fs_usage_deltas) + bch2_trans_fs_usage_apply(trans, trans->fs_usage_deltas); if (unlikely(c->gc_pos.phase)) bch2_trans_mark_gc(trans); @@ -489,31 +483,85 @@ bch2_trans_commit_write_locked(struct btree_trans *trans, do_btree_insert_one(trans, i->iter, i->k); err: if (marking) { - bch2_fs_usage_scratch_put(c, fs_usage); percpu_up_read(&c->mark_lock); } return ret; } +static noinline int maybe_do_btree_merge(struct btree_trans *trans, struct btree_iter *iter) +{ + struct btree_insert_entry *i; + struct btree *b = iter_l(iter)->b; + struct bkey_s_c old; + int u64s_delta = 0; + int ret; + + /* + * Inserting directly into interior nodes is an uncommon operation with + * various weird edge cases: also, a lot of things about + * BTREE_ITER_NODES iters need to be audited + */ + if (unlikely(btree_iter_type(iter) != BTREE_ITER_KEYS)) + return 0; + + BUG_ON(iter->level); + + trans_for_each_update2(trans, i) { + if (iter_l(i->iter)->b != b) + continue; + + old = bch2_btree_iter_peek_slot(i->iter); + ret = bkey_err(old); + if (ret) + return ret; + + u64s_delta += !bkey_deleted(&i->k->k) ? i->k->k.u64s : 0; + u64s_delta -= !bkey_deleted(old.k) ? old.k->u64s : 0; + } + + return u64s_delta <= 0 + ? (bch2_foreground_maybe_merge(trans->c, iter, iter->level, + trans->flags & ~BTREE_INSERT_NOUNLOCK) ?: -EINTR) + : 0; +} + /* * Get journal reservation, take write locks, and attempt to do btree update(s): */ static inline int do_bch2_trans_commit(struct btree_trans *trans, struct btree_insert_entry **stopped_at) { + struct bch_fs *c = trans->c; struct btree_insert_entry *i; struct btree_iter *iter; int ret; + trans_for_each_update2(trans, i) { + struct btree *b; + + BUG_ON(!btree_node_intent_locked(i->iter, i->level)); + + if (btree_iter_type(i->iter) == BTREE_ITER_CACHED) + continue; + + b = iter_l(i->iter)->b; + if (b->sib_u64s[0] < c->btree_foreground_merge_threshold || + b->sib_u64s[1] < c->btree_foreground_merge_threshold) { + ret = maybe_do_btree_merge(trans, i->iter); + if (unlikely(ret)) + return ret; + } + } + trans_for_each_update2(trans, i) - BUG_ON(!btree_node_intent_locked(i->iter, i->iter->level)); + BUG_ON(!btree_node_intent_locked(i->iter, i->level)); - ret = bch2_journal_preres_get(&trans->c->journal, + ret = bch2_journal_preres_get(&c->journal, &trans->journal_preres, trans->journal_preres_u64s, JOURNAL_RES_GET_NONBLOCK| - ((trans->flags & BTREE_INSERT_JOURNAL_RECLAIM) - ? JOURNAL_RES_GET_RECLAIM : 0)); + ((trans->flags & BTREE_INSERT_JOURNAL_RESERVED) + ? JOURNAL_RES_GET_RESERVED : 0)); if (unlikely(ret == -EAGAIN)) ret = bch2_trans_journal_preres_get_cold(trans, trans->journal_preres_u64s); @@ -522,6 +570,10 @@ static inline int do_bch2_trans_commit(struct btree_trans *trans, /* * Can't be holding any read locks when we go to take write locks: + * another thread could be holding an intent lock on the same node we + * have a read lock on, and it'll block trying to take a write lock + * (because we hold a read lock) and it could be blocking us by holding + * its own read lock (while we're trying to to take write locks). * * note - this must be done after bch2_trans_journal_preres_get_cold() * or anything else that might call bch2_trans_relock(), since that @@ -529,20 +581,25 @@ static inline int do_bch2_trans_commit(struct btree_trans *trans, */ trans_for_each_iter(trans, iter) { if (iter->nodes_locked != iter->nodes_intent_locked) { - EBUG_ON(iter->flags & BTREE_ITER_KEEP_UNTIL_COMMIT); - EBUG_ON(trans->iters_live & (1ULL << iter->idx)); - bch2_btree_iter_unlock_noinline(iter); + if (btree_iter_keep(trans, iter)) { + if (!bch2_btree_iter_upgrade(iter, 1)) { + trace_trans_restart_upgrade(trans->ip); + return -EINTR; + } + } else { + bch2_btree_iter_unlock_noinline(iter); + } } } if (IS_ENABLED(CONFIG_BCACHEFS_DEBUG)) trans_for_each_update2(trans, i) - btree_insert_entry_checks(trans, i->iter, i->k); + btree_insert_entry_checks(trans, i); bch2_btree_trans_verify_locks(trans); trans_for_each_update2(trans, i) if (!same_leaf_as_prev(trans, i)) - bch2_btree_node_lock_for_insert(trans->c, + bch2_btree_node_lock_for_insert(c, iter_l(i->iter)->b, i->iter); ret = bch2_trans_commit_write_locked(trans, stopped_at); @@ -553,32 +610,45 @@ static inline int do_bch2_trans_commit(struct btree_trans *trans, i->iter); if (!ret && trans->journal_pin) - bch2_journal_pin_add(&trans->c->journal, trans->journal_res.seq, + bch2_journal_pin_add(&c->journal, trans->journal_res.seq, trans->journal_pin, NULL); /* * Drop journal reservation after dropping write locks, since dropping * the journal reservation may kick off a journal write: */ - bch2_journal_res_put(&trans->c->journal, &trans->journal_res); + bch2_journal_res_put(&c->journal, &trans->journal_res); if (unlikely(ret)) return ret; - if (trans->flags & BTREE_INSERT_NOUNLOCK) - trans->nounlock = true; + bch2_trans_downgrade(trans); - trans_for_each_update2(trans, i) - if (btree_iter_type(i->iter) != BTREE_ITER_CACHED && - !same_leaf_as_prev(trans, i)) - bch2_foreground_maybe_merge(trans->c, i->iter, - 0, trans->flags); + return 0; +} - trans->nounlock = false; +static int journal_reclaim_wait_done(struct bch_fs *c) +{ + int ret; - bch2_trans_downgrade(trans); + ret = bch2_journal_error(&c->journal); + if (ret) + return ret; - return 0; + ret = !bch2_btree_key_cache_must_wait(c); + if (ret) + return ret; + + journal_reclaim_kick(&c->journal); + + if (mutex_trylock(&c->journal.reclaim_lock)) { + ret = bch2_journal_reclaim(&c->journal); + mutex_unlock(&c->journal.reclaim_lock); + } + + if (!ret) + ret = !bch2_btree_key_cache_must_wait(c); + return ret; } static noinline @@ -635,11 +705,9 @@ int bch2_trans_commit_error(struct btree_trans *trans, case BTREE_INSERT_NEED_MARK_REPLICAS: bch2_trans_unlock(trans); - trans_for_each_update(trans, i) { - ret = bch2_mark_bkey_replicas(c, bkey_i_to_s_c(i->k)); - if (ret) - return ret; - } + ret = bch2_replicas_delta_list_mark(c, trans->fs_usage_deltas); + if (ret) + return ret; if (bch2_trans_relock(trans)) return 0; @@ -650,6 +718,10 @@ int bch2_trans_commit_error(struct btree_trans *trans, case BTREE_INSERT_NEED_JOURNAL_RES: bch2_trans_unlock(trans); + if ((trans->flags & BTREE_INSERT_JOURNAL_RECLAIM) && + !(trans->flags & BTREE_INSERT_JOURNAL_RESERVED)) + return -EAGAIN; + ret = bch2_trans_journal_res_get(trans, JOURNAL_RES_GET_CHECK); if (ret) return ret; @@ -660,20 +732,21 @@ int bch2_trans_commit_error(struct btree_trans *trans, trace_trans_restart_journal_res_get(trans->ip); ret = -EINTR; break; - default: - BUG_ON(ret >= 0); - break; - } + case BTREE_INSERT_NEED_JOURNAL_RECLAIM: + bch2_trans_unlock(trans); - if (ret == -EINTR) { - int ret2 = bch2_btree_iter_traverse_all(trans); + wait_event(c->journal.reclaim_wait, + (ret = journal_reclaim_wait_done(c))); - if (ret2) { - trace_trans_restart_traverse(trans->ip); - return ret2; - } + if (!ret && bch2_trans_relock(trans)) + return 0; - trace_trans_restart_atomic(trans->ip); + trace_trans_restart_journal_reclaim(trans->ip); + ret = -EINTR; + break; + default: + BUG_ON(ret >= 0); + break; } return ret; @@ -698,132 +771,134 @@ bch2_trans_commit_get_rw_cold(struct btree_trans *trans) return 0; } -static void bch2_trans_update2(struct btree_trans *trans, - struct btree_iter *iter, - struct bkey_i *insert) +static void __bch2_trans_update2(struct btree_trans *trans, + struct btree_insert_entry n) { - struct btree_insert_entry *i, n = (struct btree_insert_entry) { - .iter = iter, .k = insert - }; - - btree_insert_entry_checks(trans, n.iter, n.k); + struct btree_insert_entry *i; - BUG_ON(iter->uptodate > BTREE_ITER_NEED_PEEK); + btree_insert_entry_checks(trans, &n); - EBUG_ON(trans->nr_updates2 >= trans->nr_iters); + EBUG_ON(trans->nr_updates2 >= BTREE_ITER_MAX); - iter->flags |= BTREE_ITER_KEEP_UNTIL_COMMIT; + n.iter->flags |= BTREE_ITER_KEEP_UNTIL_COMMIT; - trans_for_each_update2(trans, i) { - if (btree_iter_cmp(n.iter, i->iter) == 0) { - *i = n; - return; - } - - if (btree_iter_cmp(n.iter, i->iter) <= 0) + trans_for_each_update2(trans, i) + if (btree_insert_entry_cmp(&n, i) <= 0) break; - } - array_insert_item(trans->updates2, trans->nr_updates2, - i - trans->updates2, n); + if (i < trans->updates2 + trans->nr_updates2 && + !btree_insert_entry_cmp(&n, i)) + *i = n; + else + array_insert_item(trans->updates2, trans->nr_updates2, + i - trans->updates2, n); +} + +static void bch2_trans_update2(struct btree_trans *trans, + struct btree_iter *iter, + struct bkey_i *insert) +{ + __bch2_trans_update2(trans, (struct btree_insert_entry) { + .bkey_type = __btree_node_type(iter->level, iter->btree_id), + .btree_id = iter->btree_id, + .level = iter->level, + .iter = iter, + .k = insert, + }); } static int extent_update_to_keys(struct btree_trans *trans, - struct btree_iter *orig_iter, - struct bkey_i *insert) + struct btree_insert_entry n) { - struct btree_iter *iter; + int ret; - if (bkey_deleted(&insert->k)) + if (bkey_deleted(&n.k->k)) return 0; - iter = bch2_trans_copy_iter(trans, orig_iter); - if (IS_ERR(iter)) - return PTR_ERR(iter); + ret = bch2_extent_can_insert(trans, n.iter, n.k); + if (ret) + return ret; + + n.iter = bch2_trans_get_iter(trans, n.iter->btree_id, n.k->k.p, + BTREE_ITER_INTENT| + BTREE_ITER_NOT_EXTENTS); + n.is_extent = false; - iter->flags |= BTREE_ITER_INTENT; - __bch2_btree_iter_set_pos(iter, insert->k.p, false); - bch2_trans_update2(trans, iter, insert); - bch2_trans_iter_put(trans, iter); + __bch2_trans_update2(trans, n); + bch2_trans_iter_put(trans, n.iter); return 0; } static int extent_handle_overwrites(struct btree_trans *trans, enum btree_id btree_id, - struct bpos start, struct bpos end) + struct bkey_i *insert) { - struct btree_iter *iter = NULL, *update_iter; + struct btree_iter *iter, *update_iter; + struct bpos start = bkey_start_pos(&insert->k); struct bkey_i *update; struct bkey_s_c k; int ret = 0; - iter = bch2_trans_get_iter(trans, btree_id, start, BTREE_ITER_INTENT); - ret = PTR_ERR_OR_ZERO(iter); - if (ret) - return ret; - + iter = bch2_trans_get_iter(trans, btree_id, start, + BTREE_ITER_INTENT); k = bch2_btree_iter_peek_with_updates(iter); while (k.k && !(ret = bkey_err(k))) { - if (bkey_cmp(end, bkey_start_pos(k.k)) <= 0) + if (bkey_cmp(insert->k.p, bkey_start_pos(k.k)) <= 0) break; if (bkey_cmp(bkey_start_pos(k.k), start) < 0) { - update_iter = bch2_trans_copy_iter(trans, iter); - if ((ret = PTR_ERR_OR_ZERO(update_iter))) - goto err; - update = bch2_trans_kmalloc(trans, bkey_bytes(k.k)); if ((ret = PTR_ERR_OR_ZERO(update))) - goto err; + break; bkey_reassemble(update, k); + bch2_cut_back(start, update); - __bch2_btree_iter_set_pos(update_iter, update->k.p, false); + update_iter = bch2_trans_get_iter(trans, btree_id, update->k.p, + BTREE_ITER_NOT_EXTENTS| + BTREE_ITER_INTENT); bch2_trans_update2(trans, update_iter, update); bch2_trans_iter_put(trans, update_iter); } - if (bkey_cmp(k.k->p, end) > 0) { - update_iter = bch2_trans_copy_iter(trans, iter); - if ((ret = PTR_ERR_OR_ZERO(update_iter))) - goto err; - - update = bch2_trans_kmalloc(trans, bkey_bytes(k.k)); + if (bkey_cmp(k.k->p, insert->k.p) < 0 || + (!bkey_cmp(k.k->p, insert->k.p) && bkey_deleted(&insert->k))) { + update = bch2_trans_kmalloc(trans, sizeof(struct bkey)); if ((ret = PTR_ERR_OR_ZERO(update))) - goto err; + break; - bkey_reassemble(update, k); - bch2_cut_front(end, update); + bkey_init(&update->k); + update->k.p = k.k->p; - __bch2_btree_iter_set_pos(update_iter, update->k.p, false); + update_iter = bch2_trans_get_iter(trans, btree_id, update->k.p, + BTREE_ITER_NOT_EXTENTS| + BTREE_ITER_INTENT); bch2_trans_update2(trans, update_iter, update); bch2_trans_iter_put(trans, update_iter); - } else { - update_iter = bch2_trans_copy_iter(trans, iter); - if ((ret = PTR_ERR_OR_ZERO(update_iter))) - goto err; + } - update = bch2_trans_kmalloc(trans, sizeof(struct bkey)); + if (bkey_cmp(k.k->p, insert->k.p) > 0) { + update = bch2_trans_kmalloc(trans, bkey_bytes(k.k)); if ((ret = PTR_ERR_OR_ZERO(update))) - goto err; + break; - update->k = *k.k; - set_bkey_val_u64s(&update->k, 0); - update->k.type = KEY_TYPE_deleted; - update->k.size = 0; + bkey_reassemble(update, k); + bch2_cut_front(insert->k.p, update); - __bch2_btree_iter_set_pos(update_iter, update->k.p, false); + update_iter = bch2_trans_get_iter(trans, btree_id, update->k.p, + BTREE_ITER_NOT_EXTENTS| + BTREE_ITER_INTENT); bch2_trans_update2(trans, update_iter, update); bch2_trans_iter_put(trans, update_iter); + break; } k = bch2_btree_iter_next_with_updates(iter); } -err: - if (!IS_ERR_OR_NULL(iter)) - bch2_trans_iter_put(trans, iter); + bch2_trans_iter_put(trans, iter); + return ret; } @@ -832,13 +907,11 @@ int __bch2_trans_commit(struct btree_trans *trans) struct btree_insert_entry *i = NULL; struct btree_iter *iter; bool trans_trigger_run; - unsigned u64s; + unsigned u64s, reset_flags = 0; int ret = 0; - BUG_ON(trans->need_reset); - if (!trans->nr_updates) - goto out_noupdates; + goto out_reset; if (trans->flags & BTREE_INSERT_GC_LOCK_HELD) lockdep_assert_held(&trans->c->gc_lock); @@ -852,7 +925,7 @@ int __bch2_trans_commit(struct btree_trans *trans) unlikely(!percpu_ref_tryget(&trans->c->writes))) { ret = bch2_trans_commit_get_rw_cold(trans); if (ret) - return ret; + goto out_reset; } #ifdef CONFIG_BCACHEFS_DEBUG @@ -860,7 +933,7 @@ int __bch2_trans_commit(struct btree_trans *trans) if (btree_iter_type(i->iter) != BTREE_ITER_CACHED && !(i->trigger_flags & BTREE_TRIGGER_NORUN)) bch2_btree_key_cache_verify_clean(trans, - i->iter->btree_id, i->iter->pos); + i->btree_id, i->k->k.p); #endif /* @@ -871,24 +944,7 @@ int __bch2_trans_commit(struct btree_trans *trans) trans_trigger_run = false; trans_for_each_update(trans, i) { - if (unlikely(i->iter->uptodate > BTREE_ITER_NEED_PEEK && - (ret = bch2_btree_iter_traverse(i->iter)))) { - trace_trans_restart_traverse(trans->ip); - goto out; - } - - /* - * We're not using bch2_btree_iter_upgrade here because - * we know trans->nounlock can't be set: - */ - if (unlikely(i->iter->locks_want < 1 && - !__bch2_btree_iter_upgrade(i->iter, 1))) { - trace_trans_restart_upgrade(trans->ip); - ret = -EINTR; - goto out; - } - - if (iter_has_trans_triggers(i->iter) && + if ((BTREE_NODE_TYPE_HAS_TRANS_TRIGGERS & (1U << i->bkey_type)) && !i->trans_triggers_run) { i->trans_triggers_run = true; trans_trigger_run = true; @@ -906,33 +962,34 @@ int __bch2_trans_commit(struct btree_trans *trans) /* Turn extents updates into keys: */ trans_for_each_update(trans, i) - if (i->iter->flags & BTREE_ITER_IS_EXTENTS) { - struct bpos start = bkey_start_pos(&i->k->k); - - while (i + 1 < trans->updates + trans->nr_updates && - i[0].iter->btree_id == i[1].iter->btree_id && - !bkey_cmp(i[0].k->k.p, bkey_start_pos(&i[1].k->k))) - i++; - - ret = extent_handle_overwrites(trans, i->iter->btree_id, - start, i->k->k.p); - if (ret) + if (i->is_extent) { + ret = extent_handle_overwrites(trans, i->btree_id, i->k); + if (unlikely(ret)) goto out; } trans_for_each_update(trans, i) { - if (i->iter->flags & BTREE_ITER_IS_EXTENTS) { - ret = extent_update_to_keys(trans, i->iter, i->k); - if (ret) - goto out; - } else { - bch2_trans_update2(trans, i->iter, i->k); - } + ret = i->is_extent + ? extent_update_to_keys(trans, *i) + : (__bch2_trans_update2(trans, *i), 0); + if (unlikely(ret)) + goto out; } trans_for_each_update2(trans, i) { - BUG_ON(i->iter->uptodate > BTREE_ITER_NEED_PEEK); - BUG_ON(i->iter->locks_want < 1); + ret = bch2_btree_iter_traverse(i->iter); + if (unlikely(ret)) { + trace_trans_restart_traverse(trans->ip); + goto out; + } + + if (unlikely(!bch2_btree_iter_upgrade(i->iter, i->level + 1))) { + trace_trans_restart_upgrade(trans->ip); + ret = -EINTR; + goto out; + } + + BUG_ON(!btree_node_intent_locked(i->iter, i->level)); u64s = jset_u64s(i->k->k.u64s); if (btree_iter_type(i->iter) == BTREE_ITER_CACHED && @@ -952,20 +1009,20 @@ retry: goto err; trans_for_each_iter(trans, iter) - if ((trans->iters_live & (1ULL << iter->idx)) && - (iter->flags & BTREE_ITER_SET_POS_AFTER_COMMIT)) { - if (trans->flags & BTREE_INSERT_NOUNLOCK) - bch2_btree_iter_set_pos_same_leaf(iter, iter->pos_after_commit); - else - bch2_btree_iter_set_pos(iter, iter->pos_after_commit); - } + if (btree_iter_live(trans, iter) && + (iter->flags & BTREE_ITER_SET_POS_AFTER_COMMIT)) + bch2_btree_iter_set_pos(iter, iter->pos_after_commit); out: bch2_journal_preres_put(&trans->c->journal, &trans->journal_preres); if (likely(!(trans->flags & BTREE_INSERT_NOCHECK_RW))) percpu_ref_put(&trans->c->writes); -out_noupdates: - bch2_trans_reset(trans, !ret ? TRANS_RESET_NOTRAVERSE : 0); +out_reset: + if (!ret) + reset_flags |= TRANS_RESET_NOTRAVERSE; + if (!ret && (trans->flags & BTREE_INSERT_NOUNLOCK)) + reset_flags |= TRANS_RESET_NOUNLOCK; + bch2_trans_reset(trans, reset_flags); return ret; err: @@ -980,75 +1037,111 @@ int bch2_trans_update(struct btree_trans *trans, struct btree_iter *iter, struct bkey_i *k, enum btree_trigger_flags flags) { struct btree_insert_entry *i, n = (struct btree_insert_entry) { - .trigger_flags = flags, .iter = iter, .k = k + .trigger_flags = flags, + .bkey_type = __btree_node_type(iter->level, iter->btree_id), + .btree_id = iter->btree_id, + .level = iter->level, + .is_extent = (iter->flags & BTREE_ITER_IS_EXTENTS) != 0, + .iter = iter, + .k = k }; - EBUG_ON(bkey_cmp(iter->pos, - (iter->flags & BTREE_ITER_IS_EXTENTS) - ? bkey_start_pos(&k->k) - : k->k.p)); + BUG_ON(trans->nr_updates >= BTREE_ITER_MAX); + +#ifdef CONFIG_BCACHEFS_DEBUG + BUG_ON(bkey_cmp(iter->pos, + n.is_extent ? bkey_start_pos(&k->k) : k->k.p)); + + trans_for_each_update(trans, i) { + BUG_ON(bkey_cmp(i->iter->pos, + i->is_extent ? bkey_start_pos(&i->k->k) : i->k->k.p)); + + BUG_ON(i != trans->updates && + btree_insert_entry_cmp(i - 1, i) >= 0); + } +#endif iter->flags |= BTREE_ITER_KEEP_UNTIL_COMMIT; - if (btree_node_type_is_extents(iter->btree_id)) { + if (n.is_extent) { iter->pos_after_commit = k->k.p; iter->flags |= BTREE_ITER_SET_POS_AFTER_COMMIT; } /* - * Pending updates are kept sorted: first, find position of new update: + * Pending updates are kept sorted: first, find position of new update, + * then delete/trim any updates the new update overwrites: */ - trans_for_each_update(trans, i) - if (btree_iter_cmp(iter, i->iter) <= 0) - break; + if (!n.is_extent) { + trans_for_each_update(trans, i) + if (btree_insert_entry_cmp(&n, i) <= 0) + break; - /* - * Now delete/trim any updates the new update overwrites: - */ - if (i > trans->updates && - i[-1].iter->btree_id == iter->btree_id && - bkey_cmp(iter->pos, i[-1].k->k.p) < 0) - bch2_cut_back(n.iter->pos, i[-1].k); - - while (i < trans->updates + trans->nr_updates && - iter->btree_id == i->iter->btree_id && - bkey_cmp(n.k->k.p, i->k->k.p) >= 0) - array_remove_item(trans->updates, trans->nr_updates, - i - trans->updates); - - if (i < trans->updates + trans->nr_updates && - iter->btree_id == i->iter->btree_id && - bkey_cmp(n.k->k.p, i->iter->pos) > 0) { - /* - * When we have an extent that overwrites the start of another - * update, trimming that extent will mean the iterator's - * position has to change since the iterator position has to - * match the extent's start pos - but we don't want to change - * the iterator pos if some other code is using it, so we may - * need to clone it: - */ - if (trans->iters_live & (1ULL << i->iter->idx)) { - i->iter = bch2_trans_copy_iter(trans, i->iter); - if (IS_ERR(i->iter)) { - trans->need_reset = true; - return PTR_ERR(i->iter); + if (i < trans->updates + trans->nr_updates && + !btree_insert_entry_cmp(&n, i)) + *i = n; + else + array_insert_item(trans->updates, trans->nr_updates, + i - trans->updates, n); + } else { + trans_for_each_update(trans, i) + if (btree_insert_entry_cmp(&n, i) < 0) + break; + + while (i > trans->updates && + i[-1].btree_id == n.btree_id && + bkey_cmp(bkey_start_pos(&n.k->k), + bkey_start_pos(&i[-1].k->k)) <= 0) { + --i; + array_remove_item(trans->updates, trans->nr_updates, + i - trans->updates); + } + + if (i > trans->updates && + i[-1].btree_id == n.btree_id && + bkey_cmp(bkey_start_pos(&n.k->k), i[-1].k->k.p) < 0) + bch2_cut_back(bkey_start_pos(&n.k->k), i[-1].k); + + if (i < trans->updates + trans->nr_updates && + i->btree_id == n.btree_id && + bkey_cmp(n.k->k.p, bkey_start_pos(&i->k->k)) > 0) { + /* We don't handle splitting extents here: */ + BUG_ON(bkey_cmp(bkey_start_pos(&n.k->k), + bkey_start_pos(&i->k->k)) > 0); + + /* + * When we have an extent that overwrites the start of another + * update, trimming that extent will mean the iterator's + * position has to change since the iterator position has to + * match the extent's start pos - but we don't want to change + * the iterator pos if some other code is using it, so we may + * need to clone it: + */ + if (btree_iter_live(trans, i->iter)) { + i->iter = bch2_trans_copy_iter(trans, i->iter); + + i->iter->flags |= BTREE_ITER_KEEP_UNTIL_COMMIT; + bch2_trans_iter_put(trans, i->iter); } - i->iter->flags |= BTREE_ITER_KEEP_UNTIL_COMMIT; - bch2_trans_iter_put(trans, i->iter); + bch2_cut_front(n.k->k.p, i->k); + bch2_btree_iter_set_pos(i->iter, n.k->k.p); } - bch2_cut_front(n.k->k.p, i->k); - bch2_btree_iter_set_pos(i->iter, n.k->k.p); + array_insert_item(trans->updates, trans->nr_updates, + i - trans->updates, n); } - EBUG_ON(trans->nr_updates >= trans->nr_iters); - - array_insert_item(trans->updates, trans->nr_updates, - i - trans->updates, n); return 0; } +void bch2_trans_commit_hook(struct btree_trans *trans, + struct btree_trans_commit_hook *h) +{ + h->next = trans->hooks; + trans->hooks = h; +} + int __bch2_btree_insert(struct btree_trans *trans, enum btree_id id, struct bkey_i *k) { @@ -1057,8 +1150,6 @@ int __bch2_btree_insert(struct btree_trans *trans, iter = bch2_trans_get_iter(trans, id, bkey_start_pos(&k->k), BTREE_ITER_INTENT); - if (IS_ERR(iter)) - return PTR_ERR(iter); ret = bch2_btree_iter_traverse(iter) ?: bch2_trans_update(trans, iter, k, 0); @@ -1082,13 +1173,28 @@ int bch2_btree_insert(struct bch_fs *c, enum btree_id id, __bch2_btree_insert(&trans, id, k)); } -int bch2_btree_delete_at_range(struct btree_trans *trans, - struct btree_iter *iter, - struct bpos end, - u64 *journal_seq) +int bch2_btree_delete_at(struct btree_trans *trans, + struct btree_iter *iter, unsigned flags) +{ + struct bkey_i k; + + bkey_init(&k.k); + k.k.p = iter->pos; + + bch2_trans_update(trans, iter, &k, 0); + return bch2_trans_commit(trans, NULL, NULL, + BTREE_INSERT_NOFAIL|flags); +} + +int bch2_btree_delete_range_trans(struct btree_trans *trans, enum btree_id id, + struct bpos start, struct bpos end, + u64 *journal_seq) { + struct btree_iter *iter; struct bkey_s_c k; int ret = 0; + + iter = bch2_trans_get_iter(trans, id, start, BTREE_ITER_INTENT); retry: while ((k = bch2_btree_iter_peek(iter)).k && !(ret = bkey_err(k)) && @@ -1099,6 +1205,10 @@ retry: bkey_init(&delete.k); + /* + * This could probably be more efficient for extents: + */ + /* * For extents, iter.pos won't necessarily be the same as * bkey_start_pos(k.k) (for non extents they always will be the @@ -1138,22 +1248,8 @@ retry: goto retry; } + bch2_trans_iter_free(trans, iter); return ret; - -} - -int bch2_btree_delete_at(struct btree_trans *trans, - struct btree_iter *iter, unsigned flags) -{ - struct bkey_i k; - - bkey_init(&k.k); - k.k.p = iter->pos; - - bch2_trans_update(trans, iter, &k, 0); - return bch2_trans_commit(trans, NULL, NULL, - BTREE_INSERT_NOFAIL| - BTREE_INSERT_USE_RESERVE|flags); } /* @@ -1165,21 +1261,6 @@ int bch2_btree_delete_range(struct bch_fs *c, enum btree_id id, struct bpos start, struct bpos end, u64 *journal_seq) { - struct btree_trans trans; - struct btree_iter *iter; - int ret = 0; - - /* - * XXX: whether we need mem/more iters depends on whether this btree id - * has triggers - */ - bch2_trans_init(&trans, c, BTREE_ITER_MAX, 512); - - iter = bch2_trans_get_iter(&trans, id, start, BTREE_ITER_INTENT); - - ret = bch2_btree_delete_at_range(&trans, iter, end, journal_seq); - ret = bch2_trans_exit(&trans) ?: ret; - - BUG_ON(ret == -EINTR); - return ret; + return bch2_trans_do(c, NULL, journal_seq, 0, + bch2_btree_delete_range_trans(&trans, id, start, end, journal_seq)); }