X-Git-Url: https://git.sesse.net/?a=blobdiff_plain;f=libbcachefs%2Fbtree_iter.c;h=2bd712ebb28854b3d8b4d32ad9163fb51ec5a2dd;hb=b6740e5392a0e0a5fddc5ad4ffac0567078e114a;hp=6ee3366ed52e09598ffcad7d9ff25a77c7f45fbe;hpb=1ce6454cfe16c15bf9a36b1b58af3a7ced55d6b8;p=bcachefs-tools-debian diff --git a/libbcachefs/btree_iter.c b/libbcachefs/btree_iter.c index 6ee3366..2bd712e 100644 --- a/libbcachefs/btree_iter.c +++ b/libbcachefs/btree_iter.c @@ -5,6 +5,7 @@ #include "bkey_buf.h" #include "btree_cache.h" #include "btree_iter.h" +#include "btree_journal_iter.h" #include "btree_key_cache.h" #include "btree_locking.h" #include "btree_update.h" @@ -12,44 +13,41 @@ #include "error.h" #include "extents.h" #include "journal.h" +#include "journal_io.h" #include "replicas.h" -#include "subvolume.h" +#include "snapshot.h" +#include "trace.h" +#include #include -#include - -static void btree_trans_verify_sorted(struct btree_trans *); -static void btree_path_check_sort(struct btree_trans *, struct btree_path *, int); static inline void btree_path_list_remove(struct btree_trans *, struct btree_path *); -static inline void btree_path_list_add(struct btree_trans *, struct btree_path *, - struct btree_path *); - -static struct btree_path *btree_path_alloc(struct btree_trans *, struct btree_path *); +static inline void btree_path_list_add(struct btree_trans *, + btree_path_idx_t, btree_path_idx_t); -/* - * Unlocks before scheduling - * Note: does not revalidate iterator - */ -static inline int bch2_trans_cond_resched(struct btree_trans *trans) +static inline unsigned long btree_iter_ip_allocated(struct btree_iter *iter) { - if (need_resched() || race_fault()) { - bch2_trans_unlock(trans); - schedule(); - return bch2_trans_relock(trans) ? 0 : -EINTR; - } else { - return 0; - } +#ifdef TRACK_PATH_ALLOCATED + return iter->ip_allocated; +#else + return 0; +#endif } +static btree_path_idx_t btree_path_alloc(struct btree_trans *, btree_path_idx_t); +static void bch2_trans_srcu_lock(struct btree_trans *); + static inline int __btree_path_cmp(const struct btree_path *l, enum btree_id r_btree_id, bool r_cached, struct bpos r_pos, unsigned r_level) { + /* + * Must match lock ordering as defined by __bch2_btree_node_lock: + */ return cmp_int(l->btree_id, r_btree_id) ?: - cmp_int(l->cached, r_cached) ?: + cmp_int((int) l->cached, (int) r_cached) ?: bpos_cmp(l->pos, r_pos) ?: -cmp_int(l->level, r_level); } @@ -86,18 +84,12 @@ static inline struct bpos bkey_predecessor(struct btree_iter *iter, struct bpos return p; } -static inline bool is_btree_node(struct btree_path *path, unsigned l) -{ - return l < BTREE_MAX_DEPTH && - (unsigned long) path->l[l].b >= 128; -} - static inline struct bpos btree_iter_search_key(struct btree_iter *iter) { struct bpos pos = iter->pos; if ((iter->flags & BTREE_ITER_IS_EXTENTS) && - bkey_cmp(pos, POS_MAX)) + !bkey_eq(pos, POS_MAX)) pos = bkey_successor(iter, pos); return pos; } @@ -105,13 +97,13 @@ static inline struct bpos btree_iter_search_key(struct btree_iter *iter) static inline bool btree_path_pos_before_node(struct btree_path *path, struct btree *b) { - return bpos_cmp(path->pos, b->data->min_key) < 0; + return bpos_lt(path->pos, b->data->min_key); } static inline bool btree_path_pos_after_node(struct btree_path *path, struct btree *b) { - return bpos_cmp(b->key.k.p, path->pos) < 0; + return bpos_gt(path->pos, b->key.k.p); } static inline bool btree_path_pos_in_node(struct btree_path *path, @@ -122,435 +114,6 @@ static inline bool btree_path_pos_in_node(struct btree_path *path, !btree_path_pos_after_node(path, b); } -/* Btree node locking: */ - -void bch2_btree_node_unlock_write(struct btree_trans *trans, - struct btree_path *path, struct btree *b) -{ - bch2_btree_node_unlock_write_inlined(trans, path, b); -} - -void __bch2_btree_node_lock_write(struct btree_trans *trans, struct btree *b) -{ - struct btree_path *linked; - unsigned readers = 0; - - trans_for_each_path(trans, linked) - if (linked->l[b->c.level].b == b && - btree_node_read_locked(linked, b->c.level)) - readers++; - - /* - * Must drop our read locks before calling six_lock_write() - - * six_unlock() won't do wakeups until the reader count - * goes to 0, and it's safe because we have the node intent - * locked: - */ - atomic64_sub(__SIX_VAL(read_lock, readers), - &b->c.lock.state.counter); - btree_node_lock_type(trans->c, b, SIX_LOCK_write); - atomic64_add(__SIX_VAL(read_lock, readers), - &b->c.lock.state.counter); -} - -bool __bch2_btree_node_relock(struct btree_trans *trans, - struct btree_path *path, unsigned level) -{ - struct btree *b = btree_path_node(path, level); - int want = __btree_lock_want(path, level); - - if (!is_btree_node(path, level)) - return false; - - if (race_fault()) - return false; - - if (six_relock_type(&b->c.lock, want, path->l[level].lock_seq) || - (btree_node_lock_seq_matches(path, b, level) && - btree_node_lock_increment(trans, b, level, want))) { - mark_btree_node_locked(path, level, want); - return true; - } else { - return false; - } -} - -static bool bch2_btree_node_upgrade(struct btree_trans *trans, - struct btree_path *path, unsigned level) -{ - struct btree *b = path->l[level].b; - - if (!is_btree_node(path, level)) - return false; - - switch (btree_lock_want(path, level)) { - case BTREE_NODE_UNLOCKED: - BUG_ON(btree_node_locked(path, level)); - return true; - case BTREE_NODE_READ_LOCKED: - BUG_ON(btree_node_intent_locked(path, level)); - return bch2_btree_node_relock(trans, path, level); - case BTREE_NODE_INTENT_LOCKED: - break; - } - - if (btree_node_intent_locked(path, level)) - return true; - - if (race_fault()) - return false; - - if (btree_node_locked(path, level) - ? six_lock_tryupgrade(&b->c.lock) - : six_relock_type(&b->c.lock, SIX_LOCK_intent, path->l[level].lock_seq)) - goto success; - - if (btree_node_lock_seq_matches(path, b, level) && - btree_node_lock_increment(trans, b, level, BTREE_NODE_INTENT_LOCKED)) { - btree_node_unlock(path, level); - goto success; - } - - return false; -success: - mark_btree_node_intent_locked(path, level); - return true; -} - -static inline bool btree_path_get_locks(struct btree_trans *trans, - struct btree_path *path, - bool upgrade, unsigned long trace_ip) -{ - unsigned l = path->level; - int fail_idx = -1; - - do { - if (!btree_path_node(path, l)) - break; - - if (!(upgrade - ? bch2_btree_node_upgrade(trans, path, l) - : bch2_btree_node_relock(trans, path, l))) - fail_idx = l; - - l++; - } while (l < path->locks_want); - - /* - * When we fail to get a lock, we have to ensure that any child nodes - * can't be relocked so bch2_btree_path_traverse has to walk back up to - * the node that we failed to relock: - */ - if (fail_idx >= 0) { - __bch2_btree_path_unlock(path); - btree_path_set_dirty(path, BTREE_ITER_NEED_TRAVERSE); - - do { - path->l[fail_idx].b = BTREE_ITER_NO_NODE_GET_LOCKS; - --fail_idx; - } while (fail_idx >= 0); - } - - if (path->uptodate == BTREE_ITER_NEED_RELOCK) - path->uptodate = BTREE_ITER_UPTODATE; - - bch2_trans_verify_locks(trans); - - return path->uptodate < BTREE_ITER_NEED_RELOCK; -} - -static struct bpos btree_node_pos(struct btree_bkey_cached_common *_b, - bool cached) -{ - return !cached - ? container_of(_b, struct btree, c)->key.k.p - : container_of(_b, struct bkey_cached, c)->key.pos; -} - -/* Slowpath: */ -bool __bch2_btree_node_lock(struct btree_trans *trans, - struct btree_path *path, - struct btree *b, - struct bpos pos, unsigned level, - enum six_lock_type type, - six_lock_should_sleep_fn should_sleep_fn, void *p, - unsigned long ip) -{ - struct btree_path *linked, *deadlock_path = NULL; - u64 start_time = local_clock(); - unsigned reason = 9; - bool ret; - - /* Check if it's safe to block: */ - trans_for_each_path(trans, linked) { - if (!linked->nodes_locked) - continue; - - /* - * Can't block taking an intent lock if we have _any_ nodes read - * locked: - * - * - Our read lock blocks another thread with an intent lock on - * the same node from getting a write lock, and thus from - * dropping its intent lock - * - * - And the other thread may have multiple nodes intent locked: - * both the node we want to intent lock, and the node we - * already have read locked - deadlock: - */ - if (type == SIX_LOCK_intent && - linked->nodes_locked != linked->nodes_intent_locked) { - deadlock_path = linked; - reason = 1; - } - - if (linked->btree_id != path->btree_id) { - if (linked->btree_id > path->btree_id) { - deadlock_path = linked; - reason = 3; - } - continue; - } - - /* - * Within the same btree, cached paths come before non - * cached paths: - */ - if (linked->cached != path->cached) { - if (path->cached) { - deadlock_path = linked; - reason = 4; - } - continue; - } - - /* - * Interior nodes must be locked before their descendants: if - * another path has possible descendants locked of the node - * we're about to lock, it must have the ancestors locked too: - */ - if (level > __fls(linked->nodes_locked)) { - deadlock_path = linked; - reason = 5; - } - - /* Must lock btree nodes in key order: */ - if (btree_node_locked(linked, level) && - bpos_cmp(pos, btree_node_pos((void *) linked->l[level].b, - linked->cached)) <= 0) { - deadlock_path = linked; - reason = 7; - BUG_ON(trans->in_traverse_all); - } - } - - if (unlikely(deadlock_path)) { - trace_trans_restart_would_deadlock(trans->ip, ip, - trans->in_traverse_all, reason, - deadlock_path->btree_id, - deadlock_path->cached, - &deadlock_path->pos, - path->btree_id, - path->cached, - &pos); - btree_trans_restart(trans); - return false; - } - - if (six_trylock_type(&b->c.lock, type)) - return true; - -#ifdef CONFIG_BCACHEFS_DEBUG - trans->locking_path_idx = path->idx; - trans->locking_pos = pos; - trans->locking_btree_id = path->btree_id; - trans->locking_level = level; - trans->locking = b; -#endif - - ret = six_lock_type(&b->c.lock, type, should_sleep_fn, p) == 0; - -#ifdef CONFIG_BCACHEFS_DEBUG - trans->locking = NULL; -#endif - if (ret) - bch2_time_stats_update(&trans->c->times[lock_to_time_stat(type)], - start_time); - return ret; -} - -/* Btree iterator locking: */ - -#ifdef CONFIG_BCACHEFS_DEBUG - -static void bch2_btree_path_verify_locks(struct btree_path *path) -{ - unsigned l; - - if (!path->nodes_locked) { - BUG_ON(path->uptodate == BTREE_ITER_UPTODATE && - btree_path_node(path, path->level)); - return; - } - - for (l = 0; btree_path_node(path, l); l++) - BUG_ON(btree_lock_want(path, l) != - btree_node_locked_type(path, l)); -} - -void bch2_trans_verify_locks(struct btree_trans *trans) -{ - struct btree_path *path; - - trans_for_each_path(trans, path) - bch2_btree_path_verify_locks(path); -} -#else -static inline void bch2_btree_path_verify_locks(struct btree_path *path) {} -#endif - -/* Btree path locking: */ - -/* - * Only for btree_cache.c - only relocks intent locks - */ -bool bch2_btree_path_relock_intent(struct btree_trans *trans, - struct btree_path *path) -{ - unsigned l; - - for (l = path->level; - l < path->locks_want && btree_path_node(path, l); - l++) { - if (!bch2_btree_node_relock(trans, path, l)) { - __bch2_btree_path_unlock(path); - btree_path_set_dirty(path, BTREE_ITER_NEED_TRAVERSE); - btree_trans_restart(trans); - return false; - } - } - - return true; -} - -__flatten -static bool bch2_btree_path_relock(struct btree_trans *trans, - struct btree_path *path, unsigned long trace_ip) -{ - bool ret = btree_path_get_locks(trans, path, false, trace_ip); - - if (!ret) - btree_trans_restart(trans); - return ret; -} - -bool __bch2_btree_path_upgrade(struct btree_trans *trans, - struct btree_path *path, - unsigned new_locks_want) -{ - struct btree_path *linked; - - EBUG_ON(path->locks_want >= new_locks_want); - - path->locks_want = new_locks_want; - - if (btree_path_get_locks(trans, path, true, _THIS_IP_)) - return true; - - /* - * XXX: this is ugly - we'd prefer to not be mucking with other - * iterators in the btree_trans here. - * - * On failure to upgrade the iterator, setting iter->locks_want and - * calling get_locks() is sufficient to make bch2_btree_path_traverse() - * get the locks we want on transaction restart. - * - * But if this iterator was a clone, on transaction restart what we did - * to this iterator isn't going to be preserved. - * - * Possibly we could add an iterator field for the parent iterator when - * an iterator is a copy - for now, we'll just upgrade any other - * iterators with the same btree id. - * - * The code below used to be needed to ensure ancestor nodes get locked - * before interior nodes - now that's handled by - * bch2_btree_path_traverse_all(). - */ - trans_for_each_path(trans, linked) - if (linked != path && - linked->cached == path->cached && - linked->btree_id == path->btree_id && - linked->locks_want < new_locks_want) { - linked->locks_want = new_locks_want; - btree_path_get_locks(trans, linked, true, _THIS_IP_); - } - - return false; -} - -void __bch2_btree_path_downgrade(struct btree_path *path, - unsigned new_locks_want) -{ - unsigned l; - - EBUG_ON(path->locks_want < new_locks_want); - - path->locks_want = new_locks_want; - - while (path->nodes_locked && - (l = __fls(path->nodes_locked)) >= path->locks_want) { - if (l > path->level) { - btree_node_unlock(path, l); - } else { - if (btree_node_intent_locked(path, l)) { - six_lock_downgrade(&path->l[l].b->c.lock); - path->nodes_intent_locked ^= 1 << l; - } - break; - } - } - - bch2_btree_path_verify_locks(path); -} - -void bch2_trans_downgrade(struct btree_trans *trans) -{ - struct btree_path *path; - - trans_for_each_path(trans, path) - bch2_btree_path_downgrade(path); -} - -/* Btree transaction locking: */ - -bool bch2_trans_relock(struct btree_trans *trans) -{ - struct btree_path *path; - - if (unlikely(trans->restarted)) - return false; - - trans_for_each_path(trans, path) - if (path->should_be_locked && - !bch2_btree_path_relock(trans, path, _RET_IP_)) { - trace_trans_restart_relock(trans->ip, _RET_IP_, - path->btree_id, &path->pos); - BUG_ON(!trans->restarted); - return false; - } - return true; -} - -void bch2_trans_unlock(struct btree_trans *trans) -{ - struct btree_path *path; - - trans_for_each_path(trans, path) - __bch2_btree_path_unlock(path); - - BUG_ON(lock_class_is_held(&bch2_btree_node_lock_key)); -} - /* Btree iterator: */ #ifdef CONFIG_BCACHEFS_DEBUG @@ -566,10 +129,10 @@ static void bch2_btree_path_verify_cached(struct btree_trans *trans, ck = (void *) path->l[0].b; BUG_ON(ck->key.btree_id != path->btree_id || - bkey_cmp(ck->key.pos, path->pos)); + !bkey_eq(ck->key.pos, path->pos)); if (!locked) - btree_node_unlock(path, 0); + btree_node_unlock(trans, path, 0); } static void bch2_btree_path_verify_level(struct btree_trans *trans, @@ -579,7 +142,9 @@ static void bch2_btree_path_verify_level(struct btree_trans *trans, struct btree_node_iter tmp; bool locked; struct bkey_packed *p, *k; - char buf1[100], buf2[100], buf3[100]; + struct printbuf buf1 = PRINTBUF; + struct printbuf buf2 = PRINTBUF; + struct printbuf buf3 = PRINTBUF; const char *msg; if (!bch2_debug_check_iterators) @@ -598,7 +163,7 @@ static void bch2_btree_path_verify_level(struct btree_trans *trans, if (!btree_path_node(path, level)) return; - if (!bch2_btree_node_relock(trans, path, level)) + if (!bch2_btree_node_relock_notrace(trans, path, level)) return; BUG_ON(!btree_path_pos_in_node(path, l->b)); @@ -624,29 +189,32 @@ static void bch2_btree_path_verify_level(struct btree_trans *trans, } if (!locked) - btree_node_unlock(path, level); + btree_node_unlock(trans, path, level); return; err: - strcpy(buf2, "(none)"); - strcpy(buf3, "(none)"); - - bch2_bpos_to_text(&PBUF(buf1), path->pos); + bch2_bpos_to_text(&buf1, path->pos); if (p) { struct bkey uk = bkey_unpack_key(l->b, p); - bch2_bkey_to_text(&PBUF(buf2), &uk); + + bch2_bkey_to_text(&buf2, &uk); + } else { + prt_printf(&buf2, "(none)"); } if (k) { struct bkey uk = bkey_unpack_key(l->b, k); - bch2_bkey_to_text(&PBUF(buf3), &uk); + + bch2_bkey_to_text(&buf3, &uk); + } else { + prt_printf(&buf3, "(none)"); } panic("path should be %s key at level %u:\n" "path pos %s\n" "prev key %s\n" "cur key %s\n", - msg, level, buf1, buf2, buf3); + msg, level, buf1.buf, buf2.buf, buf3.buf); } static void bch2_btree_path_verify(struct btree_trans *trans, @@ -659,7 +227,8 @@ static void bch2_btree_path_verify(struct btree_trans *trans, for (i = 0; i < (!path->cached ? BTREE_MAX_DEPTH : 1); i++) { if (!path->l[i].b) { - BUG_ON(c->btree_roots[path->btree_id].b->c.level > i); + BUG_ON(!path->cached && + bch2_btree_id_root(c, path->btree_id)->b->c.level > i); break; } @@ -672,8 +241,9 @@ static void bch2_btree_path_verify(struct btree_trans *trans, void bch2_trans_verify_paths(struct btree_trans *trans) { struct btree_path *path; + unsigned iter; - trans_for_each_path(trans, path) + trans_for_each_path(trans, path, iter) bch2_btree_path_verify(trans, path); } @@ -683,19 +253,18 @@ static void bch2_btree_iter_verify(struct btree_iter *iter) BUG_ON(iter->btree_id >= BTREE_ID_NR); - BUG_ON(!!(iter->flags & BTREE_ITER_CACHED) != iter->path->cached); - - BUG_ON(!(iter->flags & BTREE_ITER_ALL_SNAPSHOTS) && - iter->pos.snapshot != iter->snapshot); + BUG_ON(!!(iter->flags & BTREE_ITER_CACHED) != btree_iter_path(trans, iter)->cached); BUG_ON((iter->flags & BTREE_ITER_IS_EXTENTS) && (iter->flags & BTREE_ITER_ALL_SNAPSHOTS)); BUG_ON(!(iter->flags & __BTREE_ITER_ALL_SNAPSHOTS) && (iter->flags & BTREE_ITER_ALL_SNAPSHOTS) && - !btree_type_has_snapshots(iter->btree_id)); + !btree_type_has_snapshot_field(iter->btree_id)); - bch2_btree_path_verify(trans, iter->path); + if (iter->update_path) + bch2_btree_path_verify(trans, &trans->paths[iter->update_path]); + bch2_btree_path_verify(trans, btree_iter_path(trans, iter)); } static void bch2_btree_iter_verify_entry_exit(struct btree_iter *iter) @@ -706,8 +275,8 @@ static void bch2_btree_iter_verify_entry_exit(struct btree_iter *iter) BUG_ON(!(iter->flags & BTREE_ITER_ALL_SNAPSHOTS) && iter->pos.snapshot != iter->snapshot); - BUG_ON(bkey_cmp(iter->pos, bkey_start_pos(&iter->k)) < 0 || - bkey_cmp(iter->pos, iter->k.p) > 0); + BUG_ON(bkey_lt(iter->pos, bkey_start_pos(&iter->k)) || + bkey_gt(iter->pos, iter->k.p)); } static int bch2_btree_iter_verify_ret(struct btree_iter *iter, struct bkey_s_c k) @@ -731,6 +300,7 @@ static int bch2_btree_iter_verify_ret(struct btree_iter *iter, struct bkey_s_c k k.k->p.snapshot)); bch2_trans_iter_init(trans, ©, iter->btree_id, iter->pos, + BTREE_ITER_NOPRESERVE| BTREE_ITER_ALL_SNAPSHOTS); prev = bch2_btree_iter_prev(©); if (!prev.k) @@ -740,25 +310,65 @@ static int bch2_btree_iter_verify_ret(struct btree_iter *iter, struct bkey_s_c k if (ret) goto out; - if (!bkey_cmp(prev.k->p, k.k->p) && + if (bkey_eq(prev.k->p, k.k->p) && bch2_snapshot_is_ancestor(trans->c, iter->snapshot, prev.k->p.snapshot) > 0) { - char buf1[100], buf2[200]; + struct printbuf buf1 = PRINTBUF, buf2 = PRINTBUF; - bch2_bkey_to_text(&PBUF(buf1), k.k); - bch2_bkey_to_text(&PBUF(buf2), prev.k); + bch2_bkey_to_text(&buf1, k.k); + bch2_bkey_to_text(&buf2, prev.k); panic("iter snap %u\n" "k %s\n" "prev %s\n", iter->snapshot, - buf1, buf2); + buf1.buf, buf2.buf); } out: bch2_trans_iter_exit(trans, ©); return ret; } +void bch2_assert_pos_locked(struct btree_trans *trans, enum btree_id id, + struct bpos pos, bool key_cache) +{ + struct btree_path *path; + struct trans_for_each_path_inorder_iter iter; + struct printbuf buf = PRINTBUF; + + btree_trans_sort_paths(trans); + + trans_for_each_path_inorder(trans, path, iter) { + int cmp = cmp_int(path->btree_id, id) ?: + cmp_int(path->cached, key_cache); + + if (cmp > 0) + break; + if (cmp < 0) + continue; + + if (!btree_node_locked(path, 0) || + !path->should_be_locked) + continue; + + if (!key_cache) { + if (bkey_ge(pos, path->l[0].b->data->min_key) && + bkey_le(pos, path->l[0].b->key.k.p)) + return; + } else { + if (bkey_eq(pos, path->pos)) + return; + } + } + + bch2_dump_trans_paths_updates(trans); + bch2_bpos_to_text(&buf, pos); + + panic("not locked: %s %s%s\n", + bch2_btree_id_str(id), buf.buf, + key_cache ? " cached" : ""); +} + #else static inline void bch2_btree_path_verify_level(struct btree_trans *trans, @@ -808,8 +418,9 @@ void bch2_btree_path_fix_key_modified(struct btree_trans *trans, struct bkey_packed *where) { struct btree_path *path; + unsigned i; - trans_for_each_path_with_node(trans, b, path) { + trans_for_each_path_with_node(trans, b, path, i) { __bch2_btree_path_fix_key_modified(path, b, where); bch2_btree_path_verify_level(trans, path, b->c.level); } @@ -881,7 +492,6 @@ fixup_done: if (!bch2_btree_node_iter_end(node_iter) && iter_current_key_modified && b->c.level) { - struct bset_tree *t; struct bkey_packed *k, *k2, *p; k = bch2_btree_node_iter_peek_all(node_iter, b); @@ -915,8 +525,9 @@ void bch2_btree_node_iter_fix(struct btree_trans *trans, unsigned clobber_u64s, unsigned new_u64s) { - struct bset_tree *t = bch2_bkey_to_bset(b, where); + struct bset_tree *t = bch2_bkey_to_bset_inlined(b, where); struct btree_path *linked; + unsigned i; if (node_iter != &path->l[b->c.level].iter) { __bch2_btree_node_iter_fix(path, b, node_iter, t, @@ -926,7 +537,7 @@ void bch2_btree_node_iter_fix(struct btree_trans *trans, bch2_btree_node_iter_verify(node_iter, b); } - trans_for_each_path_with_node(trans, b, linked) { + trans_for_each_path_with_node(trans, b, linked, i) { __bch2_btree_node_iter_fix(linked, b, &linked->l[b->c.level].iter, t, where, clobber_u64s, new_u64s); @@ -941,8 +552,6 @@ static inline struct bkey_s_c __btree_iter_unpack(struct bch_fs *c, struct bkey *u, struct bkey_packed *k) { - struct bkey_s_c ret; - if (unlikely(!k)) { /* * signal to bch2_btree_iter_peek_slot() that we're currently at @@ -952,19 +561,7 @@ static inline struct bkey_s_c __btree_iter_unpack(struct bch_fs *c, return bkey_s_c_null; } - ret = bkey_disassemble(l->b, k, u); - - /* - * XXX: bch2_btree_bset_insert_key() generates invalid keys when we - * overwrite extents - it sets k->type = KEY_TYPE_deleted on the key - * being overwritten but doesn't change k->size. But this is ok, because - * those keys are never written out, we just have to avoid a spurious - * assertion here: - */ - if (bch2_debug_check_bkeys && !bkey_deleted(ret.k)) - bch2_bkey_debugcheck(c, l->b, ret); - - return ret; + return bkey_disassemble(l->b, k, u); } static inline struct bkey_s_c btree_path_level_peek_all(struct bch_fs *c, @@ -975,27 +572,31 @@ static inline struct bkey_s_c btree_path_level_peek_all(struct bch_fs *c, bch2_btree_node_iter_peek_all(&l->iter, l->b)); } -static inline struct bkey_s_c btree_path_level_peek(struct bch_fs *c, +static inline struct bkey_s_c btree_path_level_peek(struct btree_trans *trans, struct btree_path *path, struct btree_path_level *l, struct bkey *u) { - struct bkey_s_c k = __btree_iter_unpack(c, l, u, + struct bkey_s_c k = __btree_iter_unpack(trans->c, l, u, bch2_btree_node_iter_peek(&l->iter, l->b)); path->pos = k.k ? k.k->p : l->b->key.k.p; + trans->paths_sorted = false; + bch2_btree_path_verify_level(trans, path, l - path->l); return k; } -static inline struct bkey_s_c btree_path_level_prev(struct bch_fs *c, +static inline struct bkey_s_c btree_path_level_prev(struct btree_trans *trans, struct btree_path *path, struct btree_path_level *l, struct bkey *u) { - struct bkey_s_c k = __btree_iter_unpack(c, l, u, + struct bkey_s_c k = __btree_iter_unpack(trans->c, l, u, bch2_btree_node_iter_prev(&l->iter, l->b)); path->pos = k.k ? k.k->p : l->b->data->min_key; + trans->paths_sorted = false; + bch2_btree_path_verify_level(trans, path, l - path->l); return k; } @@ -1018,57 +619,6 @@ static inline bool btree_path_advance_to_pos(struct btree_path *path, return true; } -/* - * Verify that iterator for parent node points to child node: - */ -static void btree_path_verify_new_node(struct btree_trans *trans, - struct btree_path *path, struct btree *b) -{ - struct btree_path_level *l; - unsigned plevel; - bool parent_locked; - struct bkey_packed *k; - - if (!IS_ENABLED(CONFIG_BCACHEFS_DEBUG)) - return; - - plevel = b->c.level + 1; - if (!btree_path_node(path, plevel)) - return; - - parent_locked = btree_node_locked(path, plevel); - - if (!bch2_btree_node_relock(trans, path, plevel)) - return; - - l = &path->l[plevel]; - k = bch2_btree_node_iter_peek_all(&l->iter, l->b); - if (!k || - bkey_deleted(k) || - bkey_cmp_left_packed(l->b, k, &b->key.k.p)) { - char buf1[100]; - char buf2[100]; - char buf3[100]; - char buf4[100]; - struct bkey uk = bkey_unpack_key(b, k); - - bch2_dump_btree_node(trans->c, l->b); - bch2_bpos_to_text(&PBUF(buf1), path->pos); - bch2_bkey_to_text(&PBUF(buf2), &uk); - bch2_bpos_to_text(&PBUF(buf3), b->data->min_key); - bch2_bpos_to_text(&PBUF(buf3), b->data->max_key); - panic("parent iter doesn't point to new node:\n" - "iter pos %s %s\n" - "iter key %s\n" - "new node %s-%s\n", - bch2_btree_ids[path->btree_id], buf1, - buf2, buf3, buf4); - } - - if (!parent_locked) - btree_node_unlock(path, plevel); -} - static inline void __btree_path_level_init(struct btree_path *path, unsigned level) { @@ -1084,47 +634,80 @@ static inline void __btree_path_level_init(struct btree_path *path, bch2_btree_node_iter_peek(&l->iter, l->b); } -static inline void btree_path_level_init(struct btree_trans *trans, - struct btree_path *path, - struct btree *b) +void bch2_btree_path_level_init(struct btree_trans *trans, + struct btree_path *path, + struct btree *b) { BUG_ON(path->cached); - btree_path_verify_new_node(trans, path, b); - EBUG_ON(!btree_path_pos_in_node(path, b)); - EBUG_ON(b->c.lock.state.seq & 1); - path->l[b->c.level].lock_seq = b->c.lock.state.seq; + path->l[b->c.level].lock_seq = six_lock_seq(&b->c.lock); path->l[b->c.level].b = b; __btree_path_level_init(path, b->c.level); } /* Btree path: fixups after btree node updates: */ +static void bch2_trans_revalidate_updates_in_node(struct btree_trans *trans, struct btree *b) +{ + struct bch_fs *c = trans->c; + struct btree_insert_entry *i; + + trans_for_each_update(trans, i) + if (!i->cached && + i->level == b->c.level && + i->btree_id == b->c.btree_id && + bpos_cmp(i->k->k.p, b->data->min_key) >= 0 && + bpos_cmp(i->k->k.p, b->data->max_key) <= 0) { + i->old_v = bch2_btree_path_peek_slot(trans->paths + i->path, &i->old_k).v; + + if (unlikely(trans->journal_replay_not_finished)) { + struct bkey_i *j_k = + bch2_journal_keys_peek_slot(c, i->btree_id, i->level, + i->k->k.p); + + if (j_k) { + i->old_k = j_k->k; + i->old_v = &j_k->v; + } + } + } +} + /* * A btree node is being replaced - update the iterator to point to the new * node: */ -void bch2_trans_node_add(struct btree_trans *trans, struct btree *b) +void bch2_trans_node_add(struct btree_trans *trans, + struct btree_path *path, + struct btree *b) { - struct btree_path *path; + struct btree_path *prev; + + BUG_ON(!btree_path_pos_in_node(path, b)); + + while ((prev = prev_btree_path(trans, path)) && + btree_path_pos_in_node(prev, b)) + path = prev; - trans_for_each_path(trans, path) - if (!path->cached && - btree_path_pos_in_node(path, b)) { + for (; + path && btree_path_pos_in_node(path, b); + path = next_btree_path(trans, path)) + if (path->uptodate == BTREE_ITER_UPTODATE && !path->cached) { enum btree_node_locked_type t = btree_lock_want(path, b->c.level); - if (path->nodes_locked && - t != BTREE_NODE_UNLOCKED) { - btree_node_unlock(path, b->c.level); - six_lock_increment(&b->c.lock, t); - mark_btree_node_locked(path, b->c.level, t); + if (t != BTREE_NODE_UNLOCKED) { + btree_node_unlock(trans, path, b->c.level); + six_lock_increment(&b->c.lock, (enum six_lock_type) t); + mark_btree_node_locked(trans, path, b->c.level, t); } - btree_path_level_init(trans, path, b); + bch2_btree_path_level_init(trans, path, b); } + + bch2_trans_revalidate_updates_in_node(trans, b); } /* @@ -1134,30 +717,26 @@ void bch2_trans_node_add(struct btree_trans *trans, struct btree *b) void bch2_trans_node_reinit_iter(struct btree_trans *trans, struct btree *b) { struct btree_path *path; + unsigned i; - trans_for_each_path_with_node(trans, b, path) + trans_for_each_path_with_node(trans, b, path, i) __btree_path_level_init(path, b->c.level); + + bch2_trans_revalidate_updates_in_node(trans, b); } /* Btree path: traverse, set_pos: */ -static int lock_root_check_fn(struct six_lock *lock, void *p) -{ - struct btree *b = container_of(lock, struct btree, c.lock); - struct btree **rootp = p; - - return b == *rootp ? 0 : -1; -} - static inline int btree_path_lock_root(struct btree_trans *trans, struct btree_path *path, unsigned depth_want, unsigned long trace_ip) { struct bch_fs *c = trans->c; - struct btree *b, **rootp = &c->btree_roots[path->btree_id].b; + struct btree *b, **rootp = &bch2_btree_id_root(c, path->btree_id)->b; enum six_lock_type lock_type; unsigned i; + int ret; EBUG_ON(path->nodes_locked); @@ -1179,26 +758,28 @@ static inline int btree_path_lock_root(struct btree_trans *trans, } lock_type = __btree_lock_want(path, path->level); - if (unlikely(!btree_node_lock(trans, path, b, SPOS_MAX, - path->level, lock_type, - lock_root_check_fn, rootp, - trace_ip))) { - if (trans->restarted) - return -EINTR; - continue; + ret = btree_node_lock(trans, path, &b->c, + path->level, lock_type, trace_ip); + if (unlikely(ret)) { + if (bch2_err_matches(ret, BCH_ERR_lock_fail_root_changed)) + continue; + if (bch2_err_matches(ret, BCH_ERR_transaction_restart)) + return ret; + BUG(); } if (likely(b == READ_ONCE(*rootp) && b->c.level == path->level && !race_fault())) { for (i = 0; i < path->level; i++) - path->l[i].b = BTREE_ITER_NO_NODE_LOCK_ROOT; + path->l[i].b = ERR_PTR(-BCH_ERR_no_btree_node_lock_root); path->l[path->level].b = b; for (i = path->level + 1; i < BTREE_MAX_DEPTH; i++) path->l[i].b = NULL; - mark_btree_node_locked(path, path->level, lock_type); - btree_path_level_init(trans, path, b); + mark_btree_node_locked(trans, path, path->level, + (enum btree_node_locked_type) lock_type); + bch2_btree_path_level_init(trans, path, b); return 0; } @@ -1214,7 +795,7 @@ static int btree_path_prefetch(struct btree_trans *trans, struct btree_path *pat struct btree_node_iter node_iter = l->iter; struct bkey_packed *k; struct bkey_buf tmp; - unsigned nr = test_bit(BCH_FS_STARTED, &c->flags) + unsigned nr = test_bit(BCH_FS_started, &c->flags) ? (path->level > 1 ? 0 : 2) : (path->level > 1 ? 1 : 16); bool was_locked = btree_node_locked(path, path->level); @@ -1222,7 +803,7 @@ static int btree_path_prefetch(struct btree_trans *trans, struct btree_path *pat bch2_bkey_buf_init(&tmp); - while (nr && !ret) { + while (nr-- && !ret) { if (!bch2_btree_node_relock(trans, path, path->level)) break; @@ -1232,12 +813,47 @@ static int btree_path_prefetch(struct btree_trans *trans, struct btree_path *pat break; bch2_bkey_buf_unpack(&tmp, c, l->b, k); - ret = bch2_btree_node_prefetch(c, trans, path, tmp.k, path->btree_id, + ret = bch2_btree_node_prefetch(trans, path, tmp.k, path->btree_id, path->level - 1); } if (!was_locked) - btree_node_unlock(path, path->level); + btree_node_unlock(trans, path, path->level); + + bch2_bkey_buf_exit(&tmp, c); + return ret; +} + +static int btree_path_prefetch_j(struct btree_trans *trans, struct btree_path *path, + struct btree_and_journal_iter *jiter) +{ + struct bch_fs *c = trans->c; + struct bkey_s_c k; + struct bkey_buf tmp; + unsigned nr = test_bit(BCH_FS_started, &c->flags) + ? (path->level > 1 ? 0 : 2) + : (path->level > 1 ? 1 : 16); + bool was_locked = btree_node_locked(path, path->level); + int ret = 0; + + bch2_bkey_buf_init(&tmp); + + while (nr-- && !ret) { + if (!bch2_btree_node_relock(trans, path, path->level)) + break; + + bch2_btree_and_journal_iter_advance(jiter); + k = bch2_btree_and_journal_iter_peek(jiter); + if (!k.k) + break; + + bch2_bkey_buf_reassemble(&tmp, c, k); + ret = bch2_btree_node_prefetch(trans, path, tmp.k, path->btree_id, + path->level - 1); + } + + if (!was_locked) + btree_node_unlock(trans, path, path->level); bch2_bkey_buf_exit(&tmp, c); return ret; @@ -1262,7 +878,31 @@ static noinline void btree_node_mem_ptr_set(struct btree_trans *trans, bp->mem_ptr = (unsigned long)b; if (!locked) - btree_node_unlock(path, plevel); + btree_node_unlock(trans, path, plevel); +} + +static noinline int btree_node_iter_and_journal_peek(struct btree_trans *trans, + struct btree_path *path, + unsigned flags, + struct bkey_buf *out) +{ + struct bch_fs *c = trans->c; + struct btree_path_level *l = path_l(path); + struct btree_and_journal_iter jiter; + struct bkey_s_c k; + int ret = 0; + + __bch2_btree_and_journal_iter_init_node_iter(&jiter, c, l->b, l->iter, path->pos); + + k = bch2_btree_and_journal_iter_peek(&jiter); + + bch2_bkey_buf_reassemble(out, c, k); + + if (flags & BTREE_ITER_PREFETCH) + ret = btree_path_prefetch_j(trans, path, &jiter); + + bch2_btree_and_journal_iter_exit(&jiter); + return ret; } static __always_inline int btree_path_down(struct btree_trans *trans, @@ -1281,27 +921,39 @@ static __always_inline int btree_path_down(struct btree_trans *trans, EBUG_ON(!btree_node_locked(path, path->level)); bch2_bkey_buf_init(&tmp); - bch2_bkey_buf_unpack(&tmp, c, l->b, - bch2_btree_node_iter_peek(&l->iter, l->b)); + + if (unlikely(trans->journal_replay_not_finished)) { + ret = btree_node_iter_and_journal_peek(trans, path, flags, &tmp); + if (ret) + goto err; + } else { + bch2_bkey_buf_unpack(&tmp, c, l->b, + bch2_btree_node_iter_peek(&l->iter, l->b)); + + if (flags & BTREE_ITER_PREFETCH) { + ret = btree_path_prefetch(trans, path); + if (ret) + goto err; + } + } b = bch2_btree_node_get(trans, path, tmp.k, level, lock_type, trace_ip); ret = PTR_ERR_OR_ZERO(b); if (unlikely(ret)) goto err; - mark_btree_node_locked(path, level, lock_type); - btree_path_level_init(trans, path, b); - - if (tmp.k->k.type == KEY_TYPE_btree_ptr_v2 && + if (likely(!trans->journal_replay_not_finished && + tmp.k->k.type == KEY_TYPE_btree_ptr_v2) && unlikely(b != btree_node_mem_ptr(tmp.k))) btree_node_mem_ptr_set(trans, path, level + 1, b); - if (flags & BTREE_ITER_PREFETCH) - ret = btree_path_prefetch(trans, path); - if (btree_node_read_locked(path, level + 1)) - btree_node_unlock(path, level + 1); + btree_node_unlock(trans, path, level + 1); + + mark_btree_node_locked(trans, path, level, + (enum btree_node_locked_type) lock_type); path->level = level; + bch2_btree_path_level_init(trans, path, b); bch2_btree_path_verify_locks(path); err: @@ -1309,139 +961,149 @@ err: return ret; } -static int btree_path_traverse_one(struct btree_trans *, struct btree_path *, - unsigned, unsigned long); -static int __btree_path_traverse_all(struct btree_trans *trans, int ret, - unsigned long trace_ip) +static int bch2_btree_path_traverse_all(struct btree_trans *trans) { struct bch_fs *c = trans->c; struct btree_path *path; - int i; + unsigned long trace_ip = _RET_IP_; + unsigned i; + int ret = 0; if (trans->in_traverse_all) - return -EINTR; + return -BCH_ERR_transaction_restart_in_traverse_all; trans->in_traverse_all = true; retry_all: - trans->restarted = false; + trans->restarted = 0; + trans->last_restarted_ip = 0; - trans_for_each_path(trans, path) + trans_for_each_path(trans, path, i) path->should_be_locked = false; - btree_trans_verify_sorted(trans); - - for (i = trans->nr_sorted - 2; i >= 0; --i) { - struct btree_path *path1 = trans->paths + trans->sorted[i]; - struct btree_path *path2 = trans->paths + trans->sorted[i + 1]; - - if (path1->btree_id == path2->btree_id && - path1->locks_want < path2->locks_want) - __bch2_btree_path_upgrade(trans, path1, path2->locks_want); - else if (!path1->locks_want && path2->locks_want) - __bch2_btree_path_upgrade(trans, path1, 1); - } + btree_trans_sort_paths(trans); bch2_trans_unlock(trans); cond_resched(); - if (unlikely(ret == -ENOMEM)) { + if (unlikely(trans->memory_allocation_failure)) { struct closure cl; closure_init_stack(&cl); do { - ret = bch2_btree_cache_cannibalize_lock(c, &cl); + ret = bch2_btree_cache_cannibalize_lock(trans, &cl); closure_sync(&cl); } while (ret); } - if (unlikely(ret == -EIO)) - goto out; - - BUG_ON(ret && ret != -EINTR); - /* Now, redo traversals in correct order: */ i = 0; while (i < trans->nr_sorted) { - path = trans->paths + trans->sorted[i]; - - EBUG_ON(!(trans->paths_allocated & (1ULL << path->idx))); - - ret = btree_path_traverse_one(trans, path, 0, _THIS_IP_); - if (ret) - goto retry_all; - - EBUG_ON(!(trans->paths_allocated & (1ULL << path->idx))); + btree_path_idx_t idx = trans->sorted[i]; - if (path->nodes_locked || - !btree_path_node(path, path->level)) + /* + * Traversing a path can cause another path to be added at about + * the same position: + */ + if (trans->paths[idx].uptodate) { + __btree_path_get(&trans->paths[idx], false); + ret = bch2_btree_path_traverse_one(trans, idx, 0, _THIS_IP_); + __btree_path_put(&trans->paths[idx], false); + + if (bch2_err_matches(ret, BCH_ERR_transaction_restart) || + bch2_err_matches(ret, ENOMEM)) + goto retry_all; + if (ret) + goto err; + } else { i++; + } } /* - * BTREE_ITER_NEED_RELOCK is ok here - if we called bch2_trans_unlock() - * and relock(), relock() won't relock since path->should_be_locked - * isn't set yet, which is all fine + * We used to assert that all paths had been traversed here + * (path->uptodate < BTREE_ITER_NEED_TRAVERSE); however, since + * path->should_be_locked is not set yet, we might have unlocked and + * then failed to relock a path - that's fine. */ - trans_for_each_path(trans, path) - BUG_ON(path->uptodate >= BTREE_ITER_NEED_TRAVERSE); -out: - bch2_btree_cache_cannibalize_unlock(c); +err: + bch2_btree_cache_cannibalize_unlock(trans); trans->in_traverse_all = false; - trace_trans_traverse_all(trans->ip, trace_ip); + trace_and_count(c, trans_traverse_all, trans, trace_ip); return ret; } -static int bch2_btree_path_traverse_all(struct btree_trans *trans) +static inline bool btree_path_check_pos_in_node(struct btree_path *path, + unsigned l, int check_pos) { - return __btree_path_traverse_all(trans, 0, _RET_IP_); + if (check_pos < 0 && btree_path_pos_before_node(path, path->l[l].b)) + return false; + if (check_pos > 0 && btree_path_pos_after_node(path, path->l[l].b)) + return false; + return true; } static inline bool btree_path_good_node(struct btree_trans *trans, struct btree_path *path, unsigned l, int check_pos) { - if (!is_btree_node(path, l) || - !bch2_btree_node_relock(trans, path, l)) - return false; + return is_btree_node(path, l) && + bch2_btree_node_relock(trans, path, l) && + btree_path_check_pos_in_node(path, l, check_pos); +} - if (check_pos < 0 && btree_path_pos_before_node(path, path->l[l].b)) - return false; - if (check_pos > 0 && btree_path_pos_after_node(path, path->l[l].b)) - return false; - return true; +static void btree_path_set_level_down(struct btree_trans *trans, + struct btree_path *path, + unsigned new_level) +{ + unsigned l; + + path->level = new_level; + + for (l = path->level + 1; l < BTREE_MAX_DEPTH; l++) + if (btree_lock_want(path, l) == BTREE_NODE_UNLOCKED) + btree_node_unlock(trans, path, l); + + btree_path_set_dirty(path, BTREE_ITER_NEED_TRAVERSE); + bch2_btree_path_verify(trans, path); } -static inline unsigned btree_path_up_until_good_node(struct btree_trans *trans, - struct btree_path *path, - int check_pos) +static noinline unsigned __btree_path_up_until_good_node(struct btree_trans *trans, + struct btree_path *path, + int check_pos) { unsigned i, l = path->level; - +again: while (btree_path_node(path, l) && - !btree_path_good_node(trans, path, l, check_pos)) { - btree_node_unlock(path, l); - path->l[l].b = BTREE_ITER_NO_NODE_UP; - l++; - } + !btree_path_good_node(trans, path, l, check_pos)) + __btree_path_set_level_up(trans, path, l++); /* If we need intent locks, take them too: */ for (i = l + 1; i < path->locks_want && btree_path_node(path, i); i++) - if (!bch2_btree_node_relock(trans, path, i)) - while (l <= i) { - btree_node_unlock(path, l); - path->l[l].b = BTREE_ITER_NO_NODE_UP; - l++; - } + if (!bch2_btree_node_relock(trans, path, i)) { + while (l <= i) + __btree_path_set_level_up(trans, path, l++); + goto again; + } return l; } +static inline unsigned btree_path_up_until_good_node(struct btree_trans *trans, + struct btree_path *path, + int check_pos) +{ + return likely(btree_node_locked(path, path->level) && + btree_path_check_pos_in_node(path, path->level, check_pos)) + ? path->level + : __btree_path_up_until_good_node(trans, path, check_pos); +} + /* * This is the main state machine for walking down the btree - walks down to a * specified depth @@ -1451,25 +1113,27 @@ static inline unsigned btree_path_up_until_good_node(struct btree_trans *trans, * On error, caller (peek_node()/peek_key()) must return NULL; the error is * stashed in the iterator and returned from bch2_trans_exit(). */ -static int btree_path_traverse_one(struct btree_trans *trans, - struct btree_path *path, - unsigned flags, - unsigned long trace_ip) +int bch2_btree_path_traverse_one(struct btree_trans *trans, + btree_path_idx_t path_idx, + unsigned flags, + unsigned long trace_ip) { + struct btree_path *path = &trans->paths[path_idx]; unsigned depth_want = path->level; - int ret = 0; + int ret = -((int) trans->restarted); - if (unlikely(trans->restarted)) { - ret = -EINTR; + if (unlikely(ret)) goto out; - } + + if (unlikely(!trans->srcu_held)) + bch2_trans_srcu_lock(trans); /* * Ensure we obey path->should_be_locked: if it's set, we can't unlock * and re-traverse the path without a transaction restart: */ if (path->should_be_locked) { - ret = bch2_btree_path_relock(trans, path, trace_ip) ? 0 : -EINTR; + ret = bch2_btree_path_relock(trans, path, trace_ip); goto out; } @@ -1478,11 +1142,16 @@ static int btree_path_traverse_one(struct btree_trans *trans, goto out; } + path = &trans->paths[path_idx]; + if (unlikely(path->level >= BTREE_MAX_DEPTH)) goto out; path->level = btree_path_up_until_good_node(trans, path, 0); + EBUG_ON(btree_path_node(path, path->level) && + !btree_node_locked(path, path->level)); + /* * Note: path->nodes[path->level] may be temporarily NULL here - that * would indicate to other code that we got to the end of the btree, @@ -1503,112 +1172,89 @@ static int btree_path_traverse_one(struct btree_trans *trans, goto out; } - __bch2_btree_path_unlock(path); + __bch2_btree_path_unlock(trans, path); path->level = depth_want; - - if (ret == -EIO) - path->l[path->level].b = - BTREE_ITER_NO_NODE_ERROR; - else - path->l[path->level].b = - BTREE_ITER_NO_NODE_DOWN; + path->l[path->level].b = ERR_PTR(ret); goto out; } } path->uptodate = BTREE_ITER_UPTODATE; out: - BUG_ON((ret == -EINTR) != !!trans->restarted); + if (bch2_err_matches(ret, BCH_ERR_transaction_restart) != !!trans->restarted) + panic("ret %s (%i) trans->restarted %s (%i)\n", + bch2_err_str(ret), ret, + bch2_err_str(trans->restarted), trans->restarted); bch2_btree_path_verify(trans, path); return ret; } -static int __btree_path_traverse_all(struct btree_trans *, int, unsigned long); - -int __must_check bch2_btree_path_traverse(struct btree_trans *trans, - struct btree_path *path, unsigned flags) -{ - if (path->uptodate < BTREE_ITER_NEED_RELOCK) - return 0; - - return bch2_trans_cond_resched(trans) ?: - btree_path_traverse_one(trans, path, flags, _RET_IP_); -} - -static void btree_path_copy(struct btree_trans *trans, struct btree_path *dst, +static inline void btree_path_copy(struct btree_trans *trans, struct btree_path *dst, struct btree_path *src) { - unsigned i; + unsigned i, offset = offsetof(struct btree_path, pos); - memcpy(&dst->pos, &src->pos, - sizeof(struct btree_path) - offsetof(struct btree_path, pos)); + memcpy((void *) dst + offset, + (void *) src + offset, + sizeof(struct btree_path) - offset); - for (i = 0; i < BTREE_MAX_DEPTH; i++) - if (btree_node_locked(dst, i)) - six_lock_increment(&dst->l[i].b->c.lock, - __btree_lock_want(dst, i)); + for (i = 0; i < BTREE_MAX_DEPTH; i++) { + unsigned t = btree_node_locked_type(dst, i); - btree_path_check_sort(trans, dst, 0); + if (t != BTREE_NODE_UNLOCKED) + six_lock_increment(&dst->l[i].b->c.lock, t); + } } -static struct btree_path *btree_path_clone(struct btree_trans *trans, struct btree_path *src, - bool intent) +static btree_path_idx_t btree_path_clone(struct btree_trans *trans, btree_path_idx_t src, + bool intent) { - struct btree_path *new = btree_path_alloc(trans, src); + btree_path_idx_t new = btree_path_alloc(trans, src); - btree_path_copy(trans, new, src); - __btree_path_get(new, intent); + btree_path_copy(trans, trans->paths + new, trans->paths + src); + __btree_path_get(trans->paths + new, intent); return new; } -inline struct btree_path * __must_check -bch2_btree_path_make_mut(struct btree_trans *trans, - struct btree_path *path, bool intent) +__flatten +btree_path_idx_t __bch2_btree_path_make_mut(struct btree_trans *trans, + btree_path_idx_t path, bool intent, unsigned long ip) { - if (path->ref > 1 || path->preserve) { - __btree_path_put(path, intent); - path = btree_path_clone(trans, path, intent); - path->preserve = false; -#ifdef CONFIG_BCACHEFS_DEBUG - path->ip_allocated = _RET_IP_; -#endif - btree_trans_verify_sorted(trans); - } - + __btree_path_put(trans->paths + path, intent); + path = btree_path_clone(trans, path, intent); + trans->paths[path].preserve = false; return path; } -static struct btree_path * __must_check -btree_path_set_pos(struct btree_trans *trans, - struct btree_path *path, struct bpos new_pos, - bool intent) +btree_path_idx_t __must_check +__bch2_btree_path_set_pos(struct btree_trans *trans, + btree_path_idx_t path_idx, struct bpos new_pos, + bool intent, unsigned long ip) { - int cmp = bpos_cmp(new_pos, path->pos); - unsigned l = path->level; + int cmp = bpos_cmp(new_pos, trans->paths[path_idx].pos); - EBUG_ON(trans->restarted); - EBUG_ON(!path->ref); + bch2_trans_verify_not_in_restart(trans); + EBUG_ON(!trans->paths[path_idx].ref); - if (!cmp) - return path; - - path = bch2_btree_path_make_mut(trans, path, intent); + path_idx = bch2_btree_path_make_mut(trans, path_idx, intent, ip); + struct btree_path *path = trans->paths + path_idx; path->pos = new_pos; - path->should_be_locked = false; - - btree_path_check_sort(trans, path, cmp); + trans->paths_sorted = false; if (unlikely(path->cached)) { - btree_node_unlock(path, 0); - path->l[0].b = BTREE_ITER_NO_NODE_CACHED; + btree_node_unlock(trans, path, 0); + path->l[0].b = ERR_PTR(-BCH_ERR_no_btree_node_up); btree_path_set_dirty(path, BTREE_ITER_NEED_TRAVERSE); goto out; } - l = btree_path_up_until_good_node(trans, path, cmp); + unsigned level = btree_path_up_until_good_node(trans, path, cmp); + + if (btree_path_node(path, level)) { + struct btree_path_level *l = &path->l[level]; - if (btree_path_node(path, l)) { + BUG_ON(!btree_node_locked(path, level)); /* * We might have to skip over many keys, or just a few: try * advancing the node iterator, and if we have to skip over too @@ -1616,155 +1262,341 @@ btree_path_set_pos(struct btree_trans *trans, * is expensive). */ if (cmp < 0 || - !btree_path_advance_to_pos(path, &path->l[l], 8)) - __btree_path_level_init(path, l); + !btree_path_advance_to_pos(path, l, 8)) + bch2_btree_node_iter_init(&l->iter, l->b, &path->pos); + + /* + * Iterators to interior nodes should always be pointed at the first non + * whiteout: + */ + if (unlikely(level)) + bch2_btree_node_iter_peek(&l->iter, l->b); } - if (l != path->level) { + if (unlikely(level != path->level)) { btree_path_set_dirty(path, BTREE_ITER_NEED_TRAVERSE); - __bch2_btree_path_unlock(path); + __bch2_btree_path_unlock(trans, path); } out: bch2_btree_path_verify(trans, path); - return path; + return path_idx; } /* Btree path: main interface: */ static struct btree_path *have_path_at_pos(struct btree_trans *trans, struct btree_path *path) { - struct btree_path *next; + struct btree_path *sib; - next = prev_btree_path(trans, path); - if (next && !btree_path_cmp(next, path)) - return next; + sib = prev_btree_path(trans, path); + if (sib && !btree_path_cmp(sib, path)) + return sib; - next = next_btree_path(trans, path); - if (next && !btree_path_cmp(next, path)) - return next; + sib = next_btree_path(trans, path); + if (sib && !btree_path_cmp(sib, path)) + return sib; return NULL; } -static bool have_node_at_pos(struct btree_trans *trans, struct btree_path *path) +static struct btree_path *have_node_at_pos(struct btree_trans *trans, struct btree_path *path) { - struct btree_path *next; + struct btree_path *sib; - next = prev_btree_path(trans, path); - if (next && path_l(next)->b == path_l(path)->b) - return true; + sib = prev_btree_path(trans, path); + if (sib && sib->level == path->level && path_l(sib)->b == path_l(path)->b) + return sib; - next = next_btree_path(trans, path); - if (next && path_l(next)->b == path_l(path)->b) - return true; + sib = next_btree_path(trans, path); + if (sib && sib->level == path->level && path_l(sib)->b == path_l(path)->b) + return sib; - return false; + return NULL; } -static inline void __bch2_path_free(struct btree_trans *trans, struct btree_path *path) +static inline void __bch2_path_free(struct btree_trans *trans, btree_path_idx_t path) { - __bch2_btree_path_unlock(path); - btree_path_list_remove(trans, path); - trans->paths_allocated &= ~(1ULL << path->idx); + __bch2_btree_path_unlock(trans, trans->paths + path); + btree_path_list_remove(trans, trans->paths + path); + __clear_bit(path, trans->paths_allocated); } -void bch2_path_put(struct btree_trans *trans, struct btree_path *path, bool intent) +void bch2_path_put(struct btree_trans *trans, btree_path_idx_t path_idx, bool intent) { - struct btree_path *dup; - - EBUG_ON(trans->paths + path->idx != path); - EBUG_ON(!path->ref); + struct btree_path *path = trans->paths + path_idx, *dup; if (!__btree_path_put(path, intent)) return; - /* - * Perhaps instead we should check for duplicate paths in traverse_all: - */ - if (path->preserve && - (dup = have_path_at_pos(trans, path))) { - dup->preserve = true; - path->preserve = false; + dup = path->preserve + ? have_path_at_pos(trans, path) + : have_node_at_pos(trans, path); + + if (!dup && !(!path->preserve && !is_btree_node(path, path->level))) + return; + + if (path->should_be_locked && + !trans->restarted && + (!dup || !bch2_btree_path_relock_norestart(trans, dup, _THIS_IP_))) + return; + + if (dup) { + dup->preserve |= path->preserve; + dup->should_be_locked |= path->should_be_locked; } - if (!path->preserve && - have_node_at_pos(trans, path)) - __bch2_path_free(trans, path); + __bch2_path_free(trans, path_idx); +} + +static void bch2_path_put_nokeep(struct btree_trans *trans, btree_path_idx_t path, + bool intent) +{ + if (!__btree_path_put(trans->paths + path, intent)) + return; + + __bch2_path_free(trans, path); +} + +void __noreturn bch2_trans_restart_error(struct btree_trans *trans, u32 restart_count) +{ + panic("trans->restart_count %u, should be %u, last restarted by %pS\n", + trans->restart_count, restart_count, + (void *) trans->last_begin_ip); +} + +void __noreturn bch2_trans_in_restart_error(struct btree_trans *trans) +{ + panic("in transaction restart: %s, last restarted by %pS\n", + bch2_err_str(trans->restarted), + (void *) trans->last_restarted_ip); } noinline __cold -void bch2_dump_trans_paths_updates(struct btree_trans *trans) +void bch2_trans_updates_to_text(struct printbuf *buf, struct btree_trans *trans) { - struct btree_path *path; struct btree_insert_entry *i; - unsigned idx; - char buf1[300], buf2[300]; - btree_trans_verify_sorted(trans); + prt_printf(buf, "transaction updates for %s journal seq %llu", + trans->fn, trans->journal_res.seq); + prt_newline(buf); + printbuf_indent_add(buf, 2); - trans_for_each_path_inorder(trans, path, idx) - printk(KERN_ERR "path: idx %u ref %u:%u%s btree %s pos %s %pS\n", - path->idx, path->ref, path->intent_ref, - path->preserve ? " preserve" : "", - bch2_btree_ids[path->btree_id], - (bch2_bpos_to_text(&PBUF(buf1), path->pos), buf1), -#ifdef CONFIG_BCACHEFS_DEBUG - (void *) path->ip_allocated -#else - NULL + trans_for_each_update(trans, i) { + struct bkey_s_c old = { &i->old_k, i->old_v }; + + prt_printf(buf, "update: btree=%s cached=%u %pS", + bch2_btree_id_str(i->btree_id), + i->cached, + (void *) i->ip_allocated); + prt_newline(buf); + + prt_printf(buf, " old "); + bch2_bkey_val_to_text(buf, trans->c, old); + prt_newline(buf); + + prt_printf(buf, " new "); + bch2_bkey_val_to_text(buf, trans->c, bkey_i_to_s_c(i->k)); + prt_newline(buf); + } + + for (struct jset_entry *e = trans->journal_entries; + e != btree_trans_journal_entries_top(trans); + e = vstruct_next(e)) + bch2_journal_entry_to_text(buf, trans->c, e); + + printbuf_indent_sub(buf, 2); +} + +noinline __cold +void bch2_dump_trans_updates(struct btree_trans *trans) +{ + struct printbuf buf = PRINTBUF; + + bch2_trans_updates_to_text(&buf, trans); + bch2_print_string_as_lines(KERN_ERR, buf.buf); + printbuf_exit(&buf); +} + +static void bch2_btree_path_to_text(struct printbuf *out, struct btree_trans *trans, btree_path_idx_t path_idx) +{ + struct btree_path *path = trans->paths + path_idx; + + prt_printf(out, "path: idx %2u ref %u:%u %c %c btree=%s l=%u pos ", + path_idx, path->ref, path->intent_ref, + path->preserve ? 'P' : ' ', + path->should_be_locked ? 'S' : ' ', + bch2_btree_id_str(path->btree_id), + path->level); + bch2_bpos_to_text(out, path->pos); + + prt_printf(out, " locks %u", path->nodes_locked); +#ifdef TRACK_PATH_ALLOCATED + prt_printf(out, " %pS", (void *) path->ip_allocated); #endif - ); + prt_newline(out); +} - trans_for_each_update(trans, i) { - struct bkey u; - struct bkey_s_c old = bch2_btree_path_peek_slot(i->path, &u); +static noinline __cold +void __bch2_trans_paths_to_text(struct printbuf *out, struct btree_trans *trans, + bool nosort) +{ + struct trans_for_each_path_inorder_iter iter; - printk(KERN_ERR "update: btree %s %pS\n old %s\n new %s", - bch2_btree_ids[i->btree_id], - (void *) i->ip_allocated, - (bch2_bkey_val_to_text(&PBUF(buf1), trans->c, old), buf1), - (bch2_bkey_val_to_text(&PBUF(buf2), trans->c, bkey_i_to_s_c(i->k)), buf2)); + if (!nosort) + btree_trans_sort_paths(trans); + + trans_for_each_path_idx_inorder(trans, iter) + bch2_btree_path_to_text(out, trans, iter.path_idx); +} + +noinline __cold +void bch2_trans_paths_to_text(struct printbuf *out, struct btree_trans *trans) +{ + __bch2_trans_paths_to_text(out, trans, false); +} + +static noinline __cold +void __bch2_dump_trans_paths_updates(struct btree_trans *trans, bool nosort) +{ + struct printbuf buf = PRINTBUF; + + __bch2_trans_paths_to_text(&buf, trans, nosort); + bch2_trans_updates_to_text(&buf, trans); + + bch2_print_string_as_lines(KERN_ERR, buf.buf); + printbuf_exit(&buf); +} + +noinline __cold +void bch2_dump_trans_paths_updates(struct btree_trans *trans) +{ + __bch2_dump_trans_paths_updates(trans, false); +} + +noinline __cold +static void bch2_trans_update_max_paths(struct btree_trans *trans) +{ + struct btree_transaction_stats *s = btree_trans_stats(trans); + struct printbuf buf = PRINTBUF; + size_t nr = bitmap_weight(trans->paths_allocated, trans->nr_paths); + + if (!s) + return; + + bch2_trans_paths_to_text(&buf, trans); + + if (!buf.allocation_failure) { + mutex_lock(&s->lock); + if (nr > s->nr_max_paths) { + s->nr_max_paths = nr; + swap(s->max_paths_text, buf.buf); + } + mutex_unlock(&s->lock); } + + printbuf_exit(&buf); + + trans->nr_paths_max = nr; } -static struct btree_path *btree_path_alloc(struct btree_trans *trans, - struct btree_path *pos) +noinline __cold +int __bch2_btree_trans_too_many_iters(struct btree_trans *trans) { - struct btree_path *path; - unsigned idx; + if (trace_trans_restart_too_many_iters_enabled()) { + struct printbuf buf = PRINTBUF; - if (unlikely(trans->paths_allocated == - ~((~0ULL << 1) << (BTREE_ITER_MAX - 1)))) { - bch2_dump_trans_paths_updates(trans); - panic("trans path oveflow\n"); + bch2_trans_paths_to_text(&buf, trans); + trace_trans_restart_too_many_iters(trans, _THIS_IP_, buf.buf); + printbuf_exit(&buf); } - idx = __ffs64(~trans->paths_allocated); - trans->paths_allocated |= 1ULL << idx; + count_event(trans->c, trans_restart_too_many_iters); + + return btree_trans_restart(trans, BCH_ERR_transaction_restart_too_many_iters); +} + +static noinline void btree_paths_realloc(struct btree_trans *trans) +{ + unsigned nr = trans->nr_paths * 2; + + void *p = kzalloc(BITS_TO_LONGS(nr) * sizeof(unsigned long) + + nr + 8 + + sizeof(struct btree_trans_paths) + + nr * sizeof(struct btree_path) + + nr * sizeof(struct btree_insert_entry), GFP_KERNEL|__GFP_NOFAIL); + + unsigned long *paths_allocated = p; + p += BITS_TO_LONGS(nr) * sizeof(unsigned long); + struct btree_path *paths = p; + p += nr * sizeof(struct btree_path); + u8 *sorted = p; + p += nr + 8; + struct btree_insert_entry *updates = p; + + *trans_paths_nr(paths) = nr; + + memcpy(paths_allocated, trans->paths_allocated, BITS_TO_LONGS(trans->nr_paths) * sizeof(unsigned long)); + memcpy(sorted, trans->sorted, trans->nr_sorted); + memcpy(paths, trans->paths, trans->nr_paths * sizeof(struct btree_path)); + memcpy(updates, trans->updates, trans->nr_paths * sizeof(struct btree_path)); + + unsigned long *old = trans->paths_allocated; + + rcu_assign_pointer(trans->paths_allocated, paths_allocated); + rcu_assign_pointer(trans->sorted, sorted); + rcu_assign_pointer(trans->paths, paths); + rcu_assign_pointer(trans->updates, updates); - path = &trans->paths[idx]; + trans->nr_paths = nr; - path->idx = idx; + if (old != trans->_paths_allocated) + kfree_rcu_mightsleep(trans->paths_allocated); +} + +static inline btree_path_idx_t btree_path_alloc(struct btree_trans *trans, + btree_path_idx_t pos) +{ + btree_path_idx_t idx = find_first_zero_bit(trans->paths_allocated, trans->nr_paths); + + if (unlikely(idx == trans->nr_paths)) + btree_paths_realloc(trans); + + /* + * Do this before marking the new path as allocated, since it won't be + * initialized yet: + */ + if (unlikely(idx > trans->nr_paths_max)) + bch2_trans_update_max_paths(trans); + + __set_bit(idx, trans->paths_allocated); + + struct btree_path *path = &trans->paths[idx]; path->ref = 0; path->intent_ref = 0; path->nodes_locked = 0; - path->nodes_intent_locked = 0; - btree_path_list_add(trans, pos, path); - return path; + btree_path_list_add(trans, pos, idx); + trans->paths_sorted = false; + return idx; } -struct btree_path *bch2_path_get(struct btree_trans *trans, bool cached, - enum btree_id btree_id, struct bpos pos, - unsigned locks_want, unsigned level, - bool intent) +btree_path_idx_t bch2_path_get(struct btree_trans *trans, + enum btree_id btree_id, struct bpos pos, + unsigned locks_want, unsigned level, + unsigned flags, unsigned long ip) { - struct btree_path *path, *path_pos = NULL; - int i; + struct btree_path *path; + bool cached = flags & BTREE_ITER_CACHED; + bool intent = flags & BTREE_ITER_INTENT; + struct trans_for_each_path_inorder_iter iter; + btree_path_idx_t path_pos = 0, path_idx; - BUG_ON(trans->restarted); + bch2_trans_verify_not_in_restart(trans); + bch2_trans_verify_locks(trans); + + btree_trans_sort_paths(trans); - trans_for_each_path_inorder(trans, path, i) { + trans_for_each_path_inorder(trans, path, iter) { if (__btree_path_cmp(path, btree_id, cached, @@ -1772,39 +1604,40 @@ struct btree_path *bch2_path_get(struct btree_trans *trans, bool cached, level) > 0) break; - path_pos = path; + path_pos = iter.path_idx; } if (path_pos && - path_pos->cached == cached && - path_pos->btree_id == btree_id && - path_pos->level == level) { - __btree_path_get(path_pos, intent); - path = btree_path_set_pos(trans, path_pos, pos, intent); - path->preserve = true; + trans->paths[path_pos].cached == cached && + trans->paths[path_pos].btree_id == btree_id && + trans->paths[path_pos].level == level) { + __btree_path_get(trans->paths + path_pos, intent); + path_idx = bch2_btree_path_set_pos(trans, path_pos, pos, intent, ip); + path = trans->paths + path_idx; } else { - path = btree_path_alloc(trans, path_pos); - path_pos = NULL; + path_idx = btree_path_alloc(trans, path_pos); + path = trans->paths + path_idx; __btree_path_get(path, intent); path->pos = pos; path->btree_id = btree_id; path->cached = cached; - path->preserve = true; path->uptodate = BTREE_ITER_NEED_TRAVERSE; path->should_be_locked = false; path->level = level; path->locks_want = locks_want; path->nodes_locked = 0; - path->nodes_intent_locked = 0; - for (i = 0; i < ARRAY_SIZE(path->l); i++) - path->l[i].b = BTREE_ITER_NO_NODE_INIT; -#ifdef CONFIG_BCACHEFS_DEBUG - path->ip_allocated = _RET_IP_; + for (unsigned i = 0; i < ARRAY_SIZE(path->l); i++) + path->l[i].b = ERR_PTR(-BCH_ERR_no_btree_node_init); +#ifdef TRACK_PATH_ALLOCATED + path->ip_allocated = ip; #endif - btree_trans_verify_sorted(trans); + trans->paths_sorted = false; } + if (!(flags & BTREE_ITER_NOPRESERVE)) + path->preserve = true; + if (path->intent_ref) locks_want = max(locks_want, level + 1); @@ -1817,42 +1650,43 @@ struct btree_path *bch2_path_get(struct btree_trans *trans, bool cached, */ locks_want = min(locks_want, BTREE_MAX_DEPTH); - if (locks_want > path->locks_want) { - path->locks_want = locks_want; - btree_path_get_locks(trans, path, true, _THIS_IP_); - } + if (locks_want > path->locks_want) + bch2_btree_path_upgrade_noupgrade_sibs(trans, path, locks_want, NULL); - return path; + return path_idx; } -inline struct bkey_s_c bch2_btree_path_peek_slot(struct btree_path *path, struct bkey *u) +struct bkey_s_c bch2_btree_path_peek_slot(struct btree_path *path, struct bkey *u) { + struct btree_path_level *l = path_l(path); + struct bkey_packed *_k; struct bkey_s_c k; - BUG_ON(path->uptodate != BTREE_ITER_UPTODATE); + if (unlikely(!l->b)) + return bkey_s_c_null; + + EBUG_ON(path->uptodate != BTREE_ITER_UPTODATE); + EBUG_ON(!btree_node_locked(path, path->level)); if (!path->cached) { - struct btree_path_level *l = path_l(path); - struct bkey_packed *_k = - bch2_btree_node_iter_peek_all(&l->iter, l->b); - + _k = bch2_btree_node_iter_peek_all(&l->iter, l->b); k = _k ? bkey_disassemble(l->b, _k, u) : bkey_s_c_null; - EBUG_ON(k.k && bkey_deleted(k.k) && bpos_cmp(k.k->p, path->pos) == 0); + EBUG_ON(k.k && bkey_deleted(k.k) && bpos_eq(k.k->p, path->pos)); - if (!k.k || bpos_cmp(path->pos, k.k->p)) + if (!k.k || !bpos_eq(path->pos, k.k->p)) goto hole; } else { struct bkey_cached *ck = (void *) path->l[0].b; - EBUG_ON(path->btree_id != ck->key.btree_id || - bkey_cmp(path->pos, ck->key.pos)); - - /* BTREE_ITER_CACHED_NOFILL? */ - if (unlikely(!ck->valid)) - goto hole; + EBUG_ON(ck && + (path->btree_id != ck->key.btree_id || + !bkey_eq(path->pos, ck->key.pos))); + if (!ck || !ck->valid) + return bkey_s_c_null; + *u = ck->k->k; k = bkey_i_to_s_c(ck->k); } @@ -1874,17 +1708,19 @@ __bch2_btree_iter_traverse(struct btree_iter *iter) int __must_check bch2_btree_iter_traverse(struct btree_iter *iter) { + struct btree_trans *trans = iter->trans; int ret; - iter->path = btree_path_set_pos(iter->trans, iter->path, + iter->path = bch2_btree_path_set_pos(trans, iter->path, btree_iter_search_key(iter), - iter->flags & BTREE_ITER_INTENT); + iter->flags & BTREE_ITER_INTENT, + btree_iter_ip_allocated(iter)); ret = bch2_btree_path_traverse(iter->trans, iter->path, iter->flags); if (ret) return ret; - iter->path->should_be_locked = true; + btree_path_set_should_be_locked(trans->paths + iter->path); return 0; } @@ -1896,26 +1732,27 @@ struct btree *bch2_btree_iter_peek_node(struct btree_iter *iter) struct btree *b = NULL; int ret; - EBUG_ON(iter->path->cached); + EBUG_ON(trans->paths[iter->path].cached); bch2_btree_iter_verify(iter); ret = bch2_btree_path_traverse(trans, iter->path, iter->flags); if (ret) goto err; - b = btree_path_node(iter->path, iter->path->level); + struct btree_path *path = btree_iter_path(trans, iter); + b = btree_path_node(path, path->level); if (!b) goto out; - BUG_ON(bpos_cmp(b->key.k.p, iter->pos) < 0); + BUG_ON(bpos_lt(b->key.k.p, iter->pos)); bkey_init(&iter->k); iter->k.p = iter->pos = b->key.k.p; - iter->path = btree_path_set_pos(trans, iter->path, b->key.k.p, - iter->flags & BTREE_ITER_INTENT); - iter->path->should_be_locked = true; - BUG_ON(iter->path->uptodate); + iter->path = bch2_btree_path_set_pos(trans, iter->path, b->key.k.p, + iter->flags & BTREE_ITER_INTENT, + btree_iter_ip_allocated(iter)); + btree_path_set_should_be_locked(btree_iter_path(trans, iter)); out: bch2_btree_iter_verify_entry_exit(iter); bch2_btree_iter_verify(iter); @@ -1926,77 +1763,82 @@ err: goto out; } +struct btree *bch2_btree_iter_peek_node_and_restart(struct btree_iter *iter) +{ + struct btree *b; + + while (b = bch2_btree_iter_peek_node(iter), + bch2_err_matches(PTR_ERR_OR_ZERO(b), BCH_ERR_transaction_restart)) + bch2_trans_begin(iter->trans); + + return b; +} + struct btree *bch2_btree_iter_next_node(struct btree_iter *iter) { struct btree_trans *trans = iter->trans; - struct btree_path *path = iter->path; struct btree *b = NULL; - unsigned l; int ret; - BUG_ON(trans->restarted); - EBUG_ON(iter->path->cached); + EBUG_ON(trans->paths[iter->path].cached); + bch2_trans_verify_not_in_restart(trans); bch2_btree_iter_verify(iter); + struct btree_path *path = btree_iter_path(trans, iter); + /* already at end? */ if (!btree_path_node(path, path->level)) return NULL; /* got to end? */ if (!btree_path_node(path, path->level + 1)) { - btree_node_unlock(path, path->level); - path->l[path->level].b = BTREE_ITER_NO_NODE_UP; - path->level++; + btree_path_set_level_up(trans, path); return NULL; } if (!bch2_btree_node_relock(trans, path, path->level + 1)) { - __bch2_btree_path_unlock(path); - path->l[path->level].b = BTREE_ITER_NO_NODE_GET_LOCKS; - path->l[path->level + 1].b = BTREE_ITER_NO_NODE_GET_LOCKS; - btree_trans_restart(trans); - ret = -EINTR; + __bch2_btree_path_unlock(trans, path); + path->l[path->level].b = ERR_PTR(-BCH_ERR_no_btree_node_relock); + path->l[path->level + 1].b = ERR_PTR(-BCH_ERR_no_btree_node_relock); + btree_path_set_dirty(path, BTREE_ITER_NEED_TRAVERSE); + trace_and_count(trans->c, trans_restart_relock_next_node, trans, _THIS_IP_, path); + ret = btree_trans_restart(trans, BCH_ERR_transaction_restart_relock); goto err; } b = btree_path_node(path, path->level + 1); - if (!bpos_cmp(iter->pos, b->key.k.p)) { - btree_node_unlock(path, path->level); - path->l[path->level].b = BTREE_ITER_NO_NODE_UP; - path->level++; + if (bpos_eq(iter->pos, b->key.k.p)) { + __btree_path_set_level_up(trans, path, path->level++); } else { /* * Haven't gotten to the end of the parent node: go back down to * the next child node */ - path = iter->path = - btree_path_set_pos(trans, path, bpos_successor(iter->pos), - iter->flags & BTREE_ITER_INTENT); + iter->path = bch2_btree_path_set_pos(trans, iter->path, + bpos_successor(iter->pos), + iter->flags & BTREE_ITER_INTENT, + btree_iter_ip_allocated(iter)); - path->level = iter->min_depth; + path = btree_iter_path(trans, iter); + btree_path_set_level_down(trans, path, iter->min_depth); - for (l = path->level + 1; l < BTREE_MAX_DEPTH; l++) - if (btree_lock_want(path, l) == BTREE_NODE_UNLOCKED) - btree_node_unlock(path, l); - - btree_path_set_dirty(path, BTREE_ITER_NEED_TRAVERSE); - bch2_btree_iter_verify(iter); - - ret = bch2_btree_path_traverse(trans, path, iter->flags); + ret = bch2_btree_path_traverse(trans, iter->path, iter->flags); if (ret) goto err; + path = btree_iter_path(trans, iter); b = path->l[path->level].b; } bkey_init(&iter->k); iter->k.p = iter->pos = b->key.k.p; - iter->path = btree_path_set_pos(trans, iter->path, b->key.k.p, - iter->flags & BTREE_ITER_INTENT); - iter->path->should_be_locked = true; - BUG_ON(iter->path->uptodate); + iter->path = bch2_btree_path_set_pos(trans, iter->path, b->key.k.p, + iter->flags & BTREE_ITER_INTENT, + btree_iter_ip_allocated(iter)); + btree_path_set_should_be_locked(btree_iter_path(trans, iter)); + EBUG_ON(btree_iter_path(trans, iter)->uptodate); out: bch2_btree_iter_verify_entry_exit(iter); bch2_btree_iter_verify(iter); @@ -2012,7 +1854,9 @@ err: inline bool bch2_btree_iter_advance(struct btree_iter *iter) { struct bpos pos = iter->k.p; - bool ret = bpos_cmp(pos, SPOS_MAX) != 0; + bool ret = !(iter->flags & BTREE_ITER_ALL_SNAPSHOTS + ? bpos_eq(pos, SPOS_MAX) + : bkey_eq(pos, SPOS_MAX)); if (ret && !(iter->flags & BTREE_ITER_IS_EXTENTS)) pos = bkey_successor(iter, pos); @@ -2023,9 +1867,9 @@ inline bool bch2_btree_iter_advance(struct btree_iter *iter) inline bool bch2_btree_iter_rewind(struct btree_iter *iter) { struct bpos pos = bkey_start_pos(&iter->k); - bool ret = (iter->flags & BTREE_ITER_ALL_SNAPSHOTS - ? bpos_cmp(pos, POS_MIN) - : bkey_cmp(pos, POS_MIN)) != 0; + bool ret = !(iter->flags & BTREE_ITER_ALL_SNAPSHOTS + ? bpos_eq(pos, POS_MIN) + : bkey_eq(pos, POS_MIN)); if (ret && !(iter->flags & BTREE_ITER_IS_EXTENTS)) pos = bkey_predecessor(iter, pos); @@ -2033,25 +1877,145 @@ inline bool bch2_btree_iter_rewind(struct btree_iter *iter) return ret; } -/** - * bch2_btree_iter_peek: returns first key greater than or equal to iterator's - * current position +static noinline +struct bkey_i *__bch2_btree_trans_peek_updates(struct btree_iter *iter) +{ + struct btree_trans *trans = iter->trans; + struct btree_insert_entry *i; + struct bkey_i *ret = NULL; + + trans_for_each_update(trans, i) { + if (i->btree_id < iter->btree_id) + continue; + if (i->btree_id > iter->btree_id) + break; + if (bpos_lt(i->k->k.p, btree_iter_path(trans, iter)->pos)) + continue; + if (i->key_cache_already_flushed) + continue; + if (!ret || bpos_lt(i->k->k.p, ret->k.p)) + ret = i->k; + } + + return ret; +} + +static inline struct bkey_i *btree_trans_peek_updates(struct btree_iter *iter) +{ + return iter->flags & BTREE_ITER_WITH_UPDATES + ? __bch2_btree_trans_peek_updates(iter) + : NULL; +} + +static struct bkey_i *bch2_btree_journal_peek(struct btree_trans *trans, + struct btree_iter *iter, + struct bpos end_pos) +{ + struct btree_path *path = btree_iter_path(trans, iter); + + return bch2_journal_keys_peek_upto(trans->c, iter->btree_id, + path->level, + path->pos, + end_pos, + &iter->journal_idx); +} + +static noinline +struct bkey_s_c btree_trans_peek_slot_journal(struct btree_trans *trans, + struct btree_iter *iter) +{ + struct btree_path *path = btree_iter_path(trans, iter); + struct bkey_i *k = bch2_btree_journal_peek(trans, iter, path->pos); + + if (k) { + iter->k = k->k; + return bkey_i_to_s_c(k); + } else { + return bkey_s_c_null; + } +} + +static noinline +struct bkey_s_c btree_trans_peek_journal(struct btree_trans *trans, + struct btree_iter *iter, + struct bkey_s_c k) +{ + struct btree_path *path = btree_iter_path(trans, iter); + struct bkey_i *next_journal = + bch2_btree_journal_peek(trans, iter, + k.k ? k.k->p : path_l(path)->b->key.k.p); + + if (next_journal) { + iter->k = next_journal->k; + k = bkey_i_to_s_c(next_journal); + } + + return k; +} + +/* + * Checks btree key cache for key at iter->pos and returns it if present, or + * bkey_s_c_null: */ -struct bkey_s_c bch2_btree_iter_peek(struct btree_iter *iter) +static noinline +struct bkey_s_c btree_trans_peek_key_cache(struct btree_iter *iter, struct bpos pos) { struct btree_trans *trans = iter->trans; - struct bpos search_key = btree_iter_search_key(iter); - struct bkey_i *next_update; + struct bch_fs *c = trans->c; + struct bkey u; struct bkey_s_c k; - int ret, cmp; + int ret; + + if ((iter->flags & BTREE_ITER_KEY_CACHE_FILL) && + bpos_eq(iter->pos, pos)) + return bkey_s_c_null; - EBUG_ON(iter->path->cached || iter->path->level); + if (!bch2_btree_key_cache_find(c, iter->btree_id, pos)) + return bkey_s_c_null; + + if (!iter->key_cache_path) + iter->key_cache_path = bch2_path_get(trans, iter->btree_id, pos, + iter->flags & BTREE_ITER_INTENT, 0, + iter->flags|BTREE_ITER_CACHED| + BTREE_ITER_CACHED_NOFILL, + _THIS_IP_); + + iter->key_cache_path = bch2_btree_path_set_pos(trans, iter->key_cache_path, pos, + iter->flags & BTREE_ITER_INTENT, + btree_iter_ip_allocated(iter)); + + ret = bch2_btree_path_traverse(trans, iter->key_cache_path, + iter->flags|BTREE_ITER_CACHED) ?: + bch2_btree_path_relock(trans, btree_iter_path(trans, iter), _THIS_IP_); + if (unlikely(ret)) + return bkey_s_c_err(ret); + + btree_path_set_should_be_locked(trans->paths + iter->key_cache_path); + + k = bch2_btree_path_peek_slot(trans->paths + iter->key_cache_path, &u); + if (k.k && !bkey_err(k)) { + iter->k = u; + k.k = &iter->k; + } + return k; +} + +static struct bkey_s_c __bch2_btree_iter_peek(struct btree_iter *iter, struct bpos search_key) +{ + struct btree_trans *trans = iter->trans; + struct bkey_i *next_update; + struct bkey_s_c k, k2; + int ret; + + EBUG_ON(btree_iter_path(trans, iter)->cached); bch2_btree_iter_verify(iter); - bch2_btree_iter_verify_entry_exit(iter); while (1) { - iter->path = btree_path_set_pos(trans, iter->path, search_key, - iter->flags & BTREE_ITER_INTENT); + struct btree_path_level *l; + + iter->path = bch2_btree_path_set_pos(trans, iter->path, search_key, + iter->flags & BTREE_ITER_INTENT, + btree_iter_ip_allocated(iter)); ret = bch2_btree_path_traverse(trans, iter->path, iter->flags); if (unlikely(ret)) { @@ -2061,91 +2025,222 @@ struct bkey_s_c bch2_btree_iter_peek(struct btree_iter *iter) goto out; } - next_update = iter->flags & BTREE_ITER_WITH_UPDATES - ? btree_trans_peek_updates(trans, iter->btree_id, search_key) - : NULL; - k = btree_path_level_peek_all(trans->c, &iter->path->l[0], &iter->k); + struct btree_path *path = btree_iter_path(trans, iter); + l = path_l(path); - /* * In the btree, deleted keys sort before non deleted: */ - if (k.k && bkey_deleted(k.k) && - (!next_update || - bpos_cmp(k.k->p, next_update->k.p) <= 0)) { - search_key = k.k->p; - continue; + if (unlikely(!l->b)) { + /* No btree nodes at requested level: */ + bch2_btree_iter_set_pos(iter, SPOS_MAX); + k = bkey_s_c_null; + goto out; + } + + btree_path_set_should_be_locked(path); + + k = btree_path_level_peek_all(trans->c, l, &iter->k); + + if (unlikely(iter->flags & BTREE_ITER_WITH_KEY_CACHE) && + k.k && + (k2 = btree_trans_peek_key_cache(iter, k.k->p)).k) { + k = k2; + ret = bkey_err(k); + if (ret) { + bch2_btree_iter_set_pos(iter, iter->pos); + goto out; + } } + if (unlikely(iter->flags & BTREE_ITER_WITH_JOURNAL)) + k = btree_trans_peek_journal(trans, iter, k); + + next_update = btree_trans_peek_updates(iter); + if (next_update && - bpos_cmp(next_update->k.p, - k.k ? k.k->p : iter->path->l[0].b->key.k.p) <= 0) { + bpos_le(next_update->k.p, + k.k ? k.k->p : l->b->key.k.p)) { iter->k = next_update->k; k = bkey_i_to_s_c(next_update); } - if (likely(k.k)) { + if (k.k && bkey_deleted(k.k)) { /* - * We can never have a key in a leaf node at POS_MAX, so - * we don't have to check these successor() calls: + * If we've got a whiteout, and it's after the search + * key, advance the search key to the whiteout instead + * of just after the whiteout - it might be a btree + * whiteout, with a real key at the same position, since + * in the btree deleted keys sort before non deleted. */ - if ((iter->flags & BTREE_ITER_FILTER_SNAPSHOTS) && - !bch2_snapshot_is_ancestor(trans->c, - iter->snapshot, - k.k->p.snapshot)) { + search_key = !bpos_eq(search_key, k.k->p) + ? k.k->p + : bpos_successor(k.k->p); + continue; + } + + if (likely(k.k)) { + break; + } else if (likely(!bpos_eq(l->b->key.k.p, SPOS_MAX))) { + /* Advance to next leaf node: */ + search_key = bpos_successor(l->b->key.k.p); + } else { + /* End of btree: */ + bch2_btree_iter_set_pos(iter, SPOS_MAX); + k = bkey_s_c_null; + goto out; + } + } +out: + bch2_btree_iter_verify(iter); + + return k; +} + +/** + * bch2_btree_iter_peek_upto() - returns first key greater than or equal to + * iterator's current position + * @iter: iterator to peek from + * @end: search limit: returns keys less than or equal to @end + * + * Returns: key if found, or an error extractable with bkey_err(). + */ +struct bkey_s_c bch2_btree_iter_peek_upto(struct btree_iter *iter, struct bpos end) +{ + struct btree_trans *trans = iter->trans; + struct bpos search_key = btree_iter_search_key(iter); + struct bkey_s_c k; + struct bpos iter_pos; + int ret; + + EBUG_ON((iter->flags & BTREE_ITER_FILTER_SNAPSHOTS) && bkey_eq(end, POS_MAX)); + + if (iter->update_path) { + bch2_path_put_nokeep(trans, iter->update_path, + iter->flags & BTREE_ITER_INTENT); + iter->update_path = 0; + } + + bch2_btree_iter_verify_entry_exit(iter); + + while (1) { + k = __bch2_btree_iter_peek(iter, search_key); + if (unlikely(!k.k)) + goto end; + if (unlikely(bkey_err(k))) + goto out_no_locked; + + /* + * iter->pos should be mononotically increasing, and always be + * equal to the key we just returned - except extents can + * straddle iter->pos: + */ + if (!(iter->flags & BTREE_ITER_IS_EXTENTS)) + iter_pos = k.k->p; + else + iter_pos = bkey_max(iter->pos, bkey_start_pos(k.k)); + + if (unlikely(!(iter->flags & BTREE_ITER_IS_EXTENTS) + ? bkey_gt(iter_pos, end) + : bkey_ge(iter_pos, end))) + goto end; + + if (iter->update_path && + !bkey_eq(trans->paths[iter->update_path].pos, k.k->p)) { + bch2_path_put_nokeep(trans, iter->update_path, + iter->flags & BTREE_ITER_INTENT); + iter->update_path = 0; + } + + if ((iter->flags & BTREE_ITER_FILTER_SNAPSHOTS) && + (iter->flags & BTREE_ITER_INTENT) && + !(iter->flags & BTREE_ITER_IS_EXTENTS) && + !iter->update_path) { + struct bpos pos = k.k->p; + + if (pos.snapshot < iter->snapshot) { search_key = bpos_successor(k.k->p); continue; } - if (bkey_whiteout(k.k) && - !(iter->flags & BTREE_ITER_ALL_SNAPSHOTS)) { - search_key = bkey_successor(iter, k.k->p); - continue; + pos.snapshot = iter->snapshot; + + /* + * advance, same as on exit for iter->path, but only up + * to snapshot + */ + __btree_path_get(trans->paths + iter->path, iter->flags & BTREE_ITER_INTENT); + iter->update_path = iter->path; + + iter->update_path = bch2_btree_path_set_pos(trans, + iter->update_path, pos, + iter->flags & BTREE_ITER_INTENT, + _THIS_IP_); + ret = bch2_btree_path_traverse(trans, iter->update_path, iter->flags); + if (unlikely(ret)) { + k = bkey_s_c_err(ret); + goto out_no_locked; } + } - break; - } else if (likely(bpos_cmp(iter->path->l[0].b->key.k.p, SPOS_MAX))) { - /* Advance to next leaf node: */ - search_key = bpos_successor(iter->path->l[0].b->key.k.p); - } else { - /* End of btree: */ - bch2_btree_iter_set_pos(iter, SPOS_MAX); - k = bkey_s_c_null; - goto out; + /* + * We can never have a key in a leaf node at POS_MAX, so + * we don't have to check these successor() calls: + */ + if ((iter->flags & BTREE_ITER_FILTER_SNAPSHOTS) && + !bch2_snapshot_is_ancestor(trans->c, + iter->snapshot, + k.k->p.snapshot)) { + search_key = bpos_successor(k.k->p); + continue; + } + + if (bkey_whiteout(k.k) && + !(iter->flags & BTREE_ITER_ALL_SNAPSHOTS)) { + search_key = bkey_successor(iter, k.k->p); + continue; } + + break; } - /* - * iter->pos should be mononotically increasing, and always be equal to - * the key we just returned - except extents can straddle iter->pos: - */ - if (!(iter->flags & BTREE_ITER_IS_EXTENTS)) - iter->pos = k.k->p; - else if (bkey_cmp(bkey_start_pos(k.k), iter->pos) > 0) - iter->pos = bkey_start_pos(k.k); + iter->pos = iter_pos; - if (iter->flags & BTREE_ITER_FILTER_SNAPSHOTS) + iter->path = bch2_btree_path_set_pos(trans, iter->path, k.k->p, + iter->flags & BTREE_ITER_INTENT, + btree_iter_ip_allocated(iter)); + + btree_path_set_should_be_locked(btree_iter_path(trans, iter)); +out_no_locked: + if (iter->update_path) { + ret = bch2_btree_path_relock(trans, trans->paths + iter->update_path, _THIS_IP_); + if (unlikely(ret)) + k = bkey_s_c_err(ret); + else + btree_path_set_should_be_locked(trans->paths + iter->update_path); + } + + if (!(iter->flags & BTREE_ITER_ALL_SNAPSHOTS)) iter->pos.snapshot = iter->snapshot; - cmp = bpos_cmp(k.k->p, iter->path->pos); - if (cmp) { - iter->path = bch2_btree_path_make_mut(trans, iter->path, - iter->flags & BTREE_ITER_INTENT); - iter->path->pos = k.k->p; - btree_path_check_sort(trans, iter->path, cmp); + ret = bch2_btree_iter_verify_ret(iter, k); + if (unlikely(ret)) { + bch2_btree_iter_set_pos(iter, iter->pos); + k = bkey_s_c_err(ret); } -out: - iter->path->should_be_locked = true; bch2_btree_iter_verify_entry_exit(iter); - bch2_btree_iter_verify(iter); - ret = bch2_btree_iter_verify_ret(iter, k); - if (unlikely(ret)) - return bkey_s_c_err(ret); return k; +end: + bch2_btree_iter_set_pos(iter, end); + k = bkey_s_c_null; + goto out_no_locked; } /** - * bch2_btree_iter_next: returns first key greater than iterator's current + * bch2_btree_iter_next() - returns first key greater than iterator's current * position + * @iter: iterator to peek from + * + * Returns: key if found, or an error extractable with bkey_err(). */ struct bkey_s_c bch2_btree_iter_next(struct btree_iter *iter) { @@ -2156,21 +2251,29 @@ struct bkey_s_c bch2_btree_iter_next(struct btree_iter *iter) } /** - * bch2_btree_iter_peek_prev: returns first key less than or equal to + * bch2_btree_iter_peek_prev() - returns first key less than or equal to * iterator's current position + * @iter: iterator to peek from + * + * Returns: key if found, or an error extractable with bkey_err(). */ struct bkey_s_c bch2_btree_iter_peek_prev(struct btree_iter *iter) { struct btree_trans *trans = iter->trans; struct bpos search_key = iter->pos; - struct btree_path *saved_path = NULL; struct bkey_s_c k; struct bkey saved_k; const struct bch_val *saved_v; + btree_path_idx_t saved_path = 0; int ret; - EBUG_ON(iter->path->cached || iter->path->level); + EBUG_ON(btree_iter_path(trans, iter)->cached || + btree_iter_path(trans, iter)->level); EBUG_ON(iter->flags & BTREE_ITER_WITH_UPDATES); + + if (iter->flags & BTREE_ITER_WITH_JOURNAL) + return bkey_s_c_err(-EIO); + bch2_btree_iter_verify(iter); bch2_btree_iter_verify_entry_exit(iter); @@ -2178,27 +2281,26 @@ struct bkey_s_c bch2_btree_iter_peek_prev(struct btree_iter *iter) search_key.snapshot = U32_MAX; while (1) { - iter->path = btree_path_set_pos(trans, iter->path, search_key, - iter->flags & BTREE_ITER_INTENT); + iter->path = bch2_btree_path_set_pos(trans, iter->path, search_key, + iter->flags & BTREE_ITER_INTENT, + btree_iter_ip_allocated(iter)); ret = bch2_btree_path_traverse(trans, iter->path, iter->flags); if (unlikely(ret)) { /* ensure that iter->k is consistent with iter->pos: */ bch2_btree_iter_set_pos(iter, iter->pos); k = bkey_s_c_err(ret); - goto out; + goto out_no_locked; } - k = btree_path_level_peek(trans->c, iter->path, - &iter->path->l[0], &iter->k); + struct btree_path *path = btree_iter_path(trans, iter); + + k = btree_path_level_peek(trans, path, &path->l[0], &iter->k); if (!k.k || ((iter->flags & BTREE_ITER_IS_EXTENTS) - ? bpos_cmp(bkey_start_pos(k.k), search_key) >= 0 - : bpos_cmp(k.k->p, search_key) > 0)) - k = btree_path_level_prev(trans->c, iter->path, - &iter->path->l[0], &iter->k); - - btree_path_check_sort(trans, iter->path, 0); + ? bpos_ge(bkey_start_pos(k.k), search_key) + : bpos_gt(k.k->p, search_key))) + k = btree_path_level_prev(trans, path, &path->l[0], &iter->k); if (likely(k.k)) { if (iter->flags & BTREE_ITER_FILTER_SNAPSHOTS) { @@ -2210,11 +2312,11 @@ struct bkey_s_c bch2_btree_iter_peek_prev(struct btree_iter *iter) * longer at the same _key_ (not pos), return * that candidate */ - if (saved_path && bkey_cmp(k.k->p, saved_k.p)) { - bch2_path_put(trans, iter->path, + if (saved_path && !bkey_eq(k.k->p, saved_k.p)) { + bch2_path_put_nokeep(trans, iter->path, iter->flags & BTREE_ITER_INTENT); iter->path = saved_path; - saved_path = NULL; + saved_path = 0; iter->k = saved_k; k.v = saved_v; goto got_key; @@ -2224,10 +2326,11 @@ struct bkey_s_c bch2_btree_iter_peek_prev(struct btree_iter *iter) iter->snapshot, k.k->p.snapshot)) { if (saved_path) - bch2_path_put(trans, saved_path, + bch2_path_put_nokeep(trans, saved_path, iter->flags & BTREE_ITER_INTENT); saved_path = btree_path_clone(trans, iter->path, iter->flags & BTREE_ITER_INTENT); + path = btree_iter_path(trans, iter); saved_k = *k.k; saved_v = k.v; } @@ -2244,30 +2347,30 @@ got_key: continue; } + btree_path_set_should_be_locked(path); break; - } else if (likely(bpos_cmp(iter->path->l[0].b->data->min_key, POS_MIN))) { + } else if (likely(!bpos_eq(path->l[0].b->data->min_key, POS_MIN))) { /* Advance to previous leaf node: */ - search_key = bpos_predecessor(iter->path->l[0].b->data->min_key); + search_key = bpos_predecessor(path->l[0].b->data->min_key); } else { /* Start of btree: */ bch2_btree_iter_set_pos(iter, POS_MIN); k = bkey_s_c_null; - goto out; + goto out_no_locked; } } - EBUG_ON(bkey_cmp(bkey_start_pos(k.k), iter->pos) > 0); + EBUG_ON(bkey_gt(bkey_start_pos(k.k), iter->pos)); /* Extents can straddle iter->pos: */ - if (bkey_cmp(k.k->p, iter->pos) < 0) + if (bkey_lt(k.k->p, iter->pos)) iter->pos = k.k->p; if (iter->flags & BTREE_ITER_FILTER_SNAPSHOTS) iter->pos.snapshot = iter->snapshot; -out: +out_no_locked: if (saved_path) - bch2_path_put(trans, saved_path, iter->flags & BTREE_ITER_INTENT); - iter->path->should_be_locked = true; + bch2_path_put_nokeep(trans, saved_path, iter->flags & BTREE_ITER_INTENT); bch2_btree_iter_verify_entry_exit(iter); bch2_btree_iter_verify(iter); @@ -2276,8 +2379,11 @@ out: } /** - * bch2_btree_iter_prev: returns first key less than iterator's current + * bch2_btree_iter_prev() - returns first key less than iterator's current * position + * @iter: iterator to peek from + * + * Returns: key if found, or an error extractable with bkey_err(). */ struct bkey_s_c bch2_btree_iter_prev(struct btree_iter *iter) { @@ -2294,9 +2400,9 @@ struct bkey_s_c bch2_btree_iter_peek_slot(struct btree_iter *iter) struct bkey_s_c k; int ret; - EBUG_ON(iter->path->level); bch2_btree_iter_verify(iter); bch2_btree_iter_verify_entry_exit(iter); + EBUG_ON(btree_iter_path(trans, iter)->level && (iter->flags & BTREE_ITER_WITH_KEY_CACHE)); /* extents can't span inode numbers: */ if ((iter->flags & BTREE_ITER_IS_EXTENTS) && @@ -2308,36 +2414,56 @@ struct bkey_s_c bch2_btree_iter_peek_slot(struct btree_iter *iter) } search_key = btree_iter_search_key(iter); - iter->path = btree_path_set_pos(trans, iter->path, search_key, - iter->flags & BTREE_ITER_INTENT); + iter->path = bch2_btree_path_set_pos(trans, iter->path, search_key, + iter->flags & BTREE_ITER_INTENT, + btree_iter_ip_allocated(iter)); ret = bch2_btree_path_traverse(trans, iter->path, iter->flags); - if (unlikely(ret)) - return bkey_s_c_err(ret); + if (unlikely(ret)) { + k = bkey_s_c_err(ret); + goto out_no_locked; + } if ((iter->flags & BTREE_ITER_CACHED) || !(iter->flags & (BTREE_ITER_IS_EXTENTS|BTREE_ITER_FILTER_SNAPSHOTS))) { struct bkey_i *next_update; - next_update = iter->flags & BTREE_ITER_WITH_UPDATES - ? btree_trans_peek_updates(trans, iter->btree_id, search_key) - : NULL; - - if (next_update && - !bpos_cmp(next_update->k.p, iter->pos)) { + if ((next_update = btree_trans_peek_updates(iter)) && + bpos_eq(next_update->k.p, iter->pos)) { iter->k = next_update->k; k = bkey_i_to_s_c(next_update); - } else { - k = bch2_btree_path_peek_slot(iter->path, &iter->k); + goto out; + } + + if (unlikely(iter->flags & BTREE_ITER_WITH_JOURNAL) && + (k = btree_trans_peek_slot_journal(trans, iter)).k) + goto out; + + if (unlikely(iter->flags & BTREE_ITER_WITH_KEY_CACHE) && + (k = btree_trans_peek_key_cache(iter, iter->pos)).k) { + if (!bkey_err(k)) + iter->k = *k.k; + /* We're not returning a key from iter->path: */ + goto out_no_locked; } + + k = bch2_btree_path_peek_slot(trans->paths + iter->path, &iter->k); + if (unlikely(!k.k)) + goto out_no_locked; } else { struct bpos next; + struct bpos end = iter->pos; + + if (iter->flags & BTREE_ITER_IS_EXTENTS) + end.offset = U64_MAX; + + EBUG_ON(btree_iter_path(trans, iter)->level); if (iter->flags & BTREE_ITER_INTENT) { struct btree_iter iter2; bch2_trans_copy_iter(&iter2, iter); - k = bch2_btree_iter_peek(&iter2); + k = bch2_btree_iter_peek_upto(&iter2, end); if (k.k && !bkey_err(k)) { iter->k = iter2.k; @@ -2347,32 +2473,38 @@ struct bkey_s_c bch2_btree_iter_peek_slot(struct btree_iter *iter) } else { struct bpos pos = iter->pos; - k = bch2_btree_iter_peek(iter); - iter->pos = pos; + k = bch2_btree_iter_peek_upto(iter, end); + if (unlikely(bkey_err(k))) + bch2_btree_iter_set_pos(iter, pos); + else + iter->pos = pos; } if (unlikely(bkey_err(k))) - return k; + goto out_no_locked; next = k.k ? bkey_start_pos(k.k) : POS_MAX; - if (bkey_cmp(iter->pos, next) < 0) { + if (bkey_lt(iter->pos, next)) { bkey_init(&iter->k); iter->k.p = iter->pos; - bch2_key_resize(&iter->k, - min_t(u64, KEY_SIZE_MAX, - (next.inode == iter->pos.inode - ? next.offset - : KEY_OFFSET_MAX) - - iter->pos.offset)); + + if (iter->flags & BTREE_ITER_IS_EXTENTS) { + bch2_key_resize(&iter->k, + min_t(u64, KEY_SIZE_MAX, + (next.inode == iter->pos.inode + ? next.offset + : KEY_OFFSET_MAX) - + iter->pos.offset)); + EBUG_ON(!iter->k.size); + } k = (struct bkey_s_c) { &iter->k, NULL }; - EBUG_ON(!k.k->size); } } - - iter->path->should_be_locked = true; - +out: + btree_path_set_should_be_locked(btree_iter_path(trans, iter)); +out_no_locked: bch2_btree_iter_verify_entry_exit(iter); bch2_btree_iter_verify(iter); ret = bch2_btree_iter_verify_ret(iter, k); @@ -2398,76 +2530,99 @@ struct bkey_s_c bch2_btree_iter_prev_slot(struct btree_iter *iter) return bch2_btree_iter_peek_slot(iter); } -/* new transactional stuff: */ - -static inline void btree_path_verify_sorted_ref(struct btree_trans *trans, - struct btree_path *path) +struct bkey_s_c bch2_btree_iter_peek_and_restart_outlined(struct btree_iter *iter) { - EBUG_ON(path->sorted_idx >= trans->nr_sorted); - EBUG_ON(trans->sorted[path->sorted_idx] != path->idx); - EBUG_ON(!(trans->paths_allocated & (1ULL << path->idx))); + struct bkey_s_c k; + + while (btree_trans_too_many_iters(iter->trans) || + (k = bch2_btree_iter_peek_type(iter, iter->flags), + bch2_err_matches(bkey_err(k), BCH_ERR_transaction_restart))) + bch2_trans_begin(iter->trans); + + return k; } -static inline void btree_trans_verify_sorted_refs(struct btree_trans *trans) -{ +/* new transactional stuff: */ + #ifdef CONFIG_BCACHEFS_DEBUG +static void btree_trans_verify_sorted_refs(struct btree_trans *trans) +{ + struct btree_path *path; unsigned i; - for (i = 0; i < trans->nr_sorted; i++) - btree_path_verify_sorted_ref(trans, trans->paths + trans->sorted[i]); -#endif + BUG_ON(trans->nr_sorted != bitmap_weight(trans->paths_allocated, trans->nr_paths) - 1); + + trans_for_each_path(trans, path, i) { + BUG_ON(path->sorted_idx >= trans->nr_sorted); + BUG_ON(trans->sorted[path->sorted_idx] != i); + } + + for (i = 0; i < trans->nr_sorted; i++) { + unsigned idx = trans->sorted[i]; + + BUG_ON(!test_bit(idx, trans->paths_allocated)); + BUG_ON(trans->paths[idx].sorted_idx != i); + } } static void btree_trans_verify_sorted(struct btree_trans *trans) { -#ifdef CONFIG_BCACHEFS_DEBUG struct btree_path *path, *prev = NULL; - unsigned i; + struct trans_for_each_path_inorder_iter iter; + + if (!bch2_debug_check_iterators) + return; - trans_for_each_path_inorder(trans, path, i) { - BUG_ON(prev && btree_path_cmp(prev, path) > 0); + trans_for_each_path_inorder(trans, path, iter) { + if (prev && btree_path_cmp(prev, path) > 0) { + __bch2_dump_trans_paths_updates(trans, true); + panic("trans paths out of order!\n"); + } prev = path; } -#endif } +#else +static inline void btree_trans_verify_sorted_refs(struct btree_trans *trans) {} +static inline void btree_trans_verify_sorted(struct btree_trans *trans) {} +#endif -static inline void btree_path_swap(struct btree_trans *trans, - struct btree_path *l, struct btree_path *r) +void __bch2_btree_trans_sort_paths(struct btree_trans *trans) { - swap(l->sorted_idx, r->sorted_idx); - swap(trans->sorted[l->sorted_idx], - trans->sorted[r->sorted_idx]); - - btree_path_verify_sorted_ref(trans, l); - btree_path_verify_sorted_ref(trans, r); -} + int i, l = 0, r = trans->nr_sorted, inc = 1; + bool swapped; -static void btree_path_check_sort(struct btree_trans *trans, struct btree_path *path, - int cmp) -{ - struct btree_path *n; + btree_trans_verify_sorted_refs(trans); - if (cmp <= 0) { - n = prev_btree_path(trans, path); - if (n && btree_path_cmp(n, path) > 0) { - do { - btree_path_swap(trans, n, path); - n = prev_btree_path(trans, path); - } while (n && btree_path_cmp(n, path) > 0); + if (trans->paths_sorted) + goto out; - goto out; + /* + * Cocktail shaker sort: this is efficient because iterators will be + * mostly sorted. + */ + do { + swapped = false; + + for (i = inc > 0 ? l : r - 2; + i + 1 < r && i >= l; + i += inc) { + if (btree_path_cmp(trans->paths + trans->sorted[i], + trans->paths + trans->sorted[i + 1]) > 0) { + swap(trans->sorted[i], trans->sorted[i + 1]); + trans->paths[trans->sorted[i]].sorted_idx = i; + trans->paths[trans->sorted[i + 1]].sorted_idx = i + 1; + swapped = true; + } } - } - if (cmp >= 0) { - n = next_btree_path(trans, path); - if (n && btree_path_cmp(path, n) > 0) { - do { - btree_path_swap(trans, path, n); - n = next_btree_path(trans, path); - } while (n && btree_path_cmp(path, n) > 0); - } - } + if (inc > 0) + --r; + else + l++; + inc = -inc; + } while (swapped); + + trans->paths_sorted = true; out: btree_trans_verify_sorted(trans); } @@ -2478,28 +2633,38 @@ static inline void btree_path_list_remove(struct btree_trans *trans, unsigned i; EBUG_ON(path->sorted_idx >= trans->nr_sorted); - +#ifdef CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS + trans->nr_sorted--; + memmove_u64s_down_small(trans->sorted + path->sorted_idx, + trans->sorted + path->sorted_idx + 1, + DIV_ROUND_UP(trans->nr_sorted - path->sorted_idx, 8)); +#else array_remove_item(trans->sorted, trans->nr_sorted, path->sorted_idx); - +#endif for (i = path->sorted_idx; i < trans->nr_sorted; i++) trans->paths[trans->sorted[i]].sorted_idx = i; path->sorted_idx = U8_MAX; - - btree_trans_verify_sorted_refs(trans); } static inline void btree_path_list_add(struct btree_trans *trans, - struct btree_path *pos, - struct btree_path *path) + btree_path_idx_t pos, + btree_path_idx_t path_idx) { + struct btree_path *path = trans->paths + path_idx; unsigned i; - btree_trans_verify_sorted_refs(trans); - - path->sorted_idx = pos ? pos->sorted_idx + 1 : 0; + path->sorted_idx = pos ? trans->paths[pos].sorted_idx + 1 : trans->nr_sorted; - array_insert_item(trans->sorted, trans->nr_sorted, path->sorted_idx, path->idx); +#ifdef CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS + memmove_u64s_up_small(trans->sorted + path->sorted_idx + 1, + trans->sorted + path->sorted_idx, + DIV_ROUND_UP(trans->nr_sorted - path->sorted_idx, 8)); + trans->nr_sorted++; + trans->sorted[path->sorted_idx] = path_idx; +#else + array_insert_item(trans->sorted, trans->nr_sorted, path->sorted_idx, path_idx); +#endif for (i = path->sorted_idx; i < trans->nr_sorted; i++) trans->paths[trans->sorted[i]].sorted_idx = i; @@ -2509,60 +2674,29 @@ static inline void btree_path_list_add(struct btree_trans *trans, void bch2_trans_iter_exit(struct btree_trans *trans, struct btree_iter *iter) { + if (iter->update_path) + bch2_path_put_nokeep(trans, iter->update_path, + iter->flags & BTREE_ITER_INTENT); if (iter->path) bch2_path_put(trans, iter->path, iter->flags & BTREE_ITER_INTENT); - iter->path = NULL; + if (iter->key_cache_path) + bch2_path_put(trans, iter->key_cache_path, + iter->flags & BTREE_ITER_INTENT); + iter->path = 0; + iter->update_path = 0; + iter->key_cache_path = 0; + iter->trans = NULL; } -static void __bch2_trans_iter_init(struct btree_trans *trans, - struct btree_iter *iter, - unsigned btree_id, struct bpos pos, - unsigned locks_want, - unsigned depth, - unsigned flags) -{ - EBUG_ON(trans->restarted); - - if (!(flags & (BTREE_ITER_ALL_SNAPSHOTS|BTREE_ITER_NOT_EXTENTS)) && - btree_node_type_is_extents(btree_id)) - flags |= BTREE_ITER_IS_EXTENTS; - - if (!(flags & __BTREE_ITER_ALL_SNAPSHOTS) && - !btree_type_has_snapshots(btree_id)) - flags &= ~BTREE_ITER_ALL_SNAPSHOTS; - - if (!(flags & BTREE_ITER_ALL_SNAPSHOTS) && - btree_type_has_snapshots(btree_id)) - flags |= BTREE_ITER_FILTER_SNAPSHOTS; - - iter->trans = trans; - iter->path = NULL; - iter->btree_id = btree_id; - iter->min_depth = depth; - iter->flags = flags; - iter->snapshot = pos.snapshot; - iter->pos = pos; - iter->k.type = KEY_TYPE_deleted; - iter->k.p = pos; - iter->k.size = 0; - - iter->path = bch2_path_get(trans, - flags & BTREE_ITER_CACHED, - btree_id, - iter->pos, - locks_want, - depth, - flags & BTREE_ITER_INTENT); -} - -void bch2_trans_iter_init(struct btree_trans *trans, +void bch2_trans_iter_init_outlined(struct btree_trans *trans, struct btree_iter *iter, - unsigned btree_id, struct bpos pos, + enum btree_id btree_id, struct bpos pos, unsigned flags) { - __bch2_trans_iter_init(trans, iter, btree_id, pos, - 0, 0, flags); + bch2_trans_iter_init_common(trans, iter, btree_id, pos, 0, 0, + bch2_btree_iter_flags(trans, btree_id, flags), + _RET_IP_); } void bch2_trans_node_iter_init(struct btree_trans *trans, @@ -2573,53 +2707,78 @@ void bch2_trans_node_iter_init(struct btree_trans *trans, unsigned depth, unsigned flags) { - __bch2_trans_iter_init(trans, iter, btree_id, pos, locks_want, depth, - BTREE_ITER_NOT_EXTENTS| - __BTREE_ITER_ALL_SNAPSHOTS| - BTREE_ITER_ALL_SNAPSHOTS| - flags); - BUG_ON(iter->path->locks_want < min(locks_want, BTREE_MAX_DEPTH)); - BUG_ON(iter->path->level != depth); - BUG_ON(iter->min_depth != depth); + flags |= BTREE_ITER_NOT_EXTENTS; + flags |= __BTREE_ITER_ALL_SNAPSHOTS; + flags |= BTREE_ITER_ALL_SNAPSHOTS; + + bch2_trans_iter_init_common(trans, iter, btree_id, pos, locks_want, depth, + __bch2_btree_iter_flags(trans, btree_id, flags), + _RET_IP_); + + iter->min_depth = depth; + + struct btree_path *path = btree_iter_path(trans, iter); + BUG_ON(path->locks_want < min(locks_want, BTREE_MAX_DEPTH)); + BUG_ON(path->level != depth); + BUG_ON(iter->min_depth != depth); } void bch2_trans_copy_iter(struct btree_iter *dst, struct btree_iter *src) { + struct btree_trans *trans = src->trans; + *dst = *src; if (src->path) - __btree_path_get(src->path, src->flags & BTREE_ITER_INTENT); + __btree_path_get(trans->paths + src->path, src->flags & BTREE_ITER_INTENT); + if (src->update_path) + __btree_path_get(trans->paths + src->update_path, src->flags & BTREE_ITER_INTENT); + dst->key_cache_path = 0; } -void *bch2_trans_kmalloc(struct btree_trans *trans, size_t size) +void *__bch2_trans_kmalloc(struct btree_trans *trans, size_t size) { - size_t new_top = trans->mem_top + size; + struct bch_fs *c = trans->c; + unsigned new_top = trans->mem_top + size; + unsigned old_bytes = trans->mem_bytes; + unsigned new_bytes = roundup_pow_of_two(new_top); + int ret; + void *new_mem; void *p; - if (new_top > trans->mem_bytes) { - size_t old_bytes = trans->mem_bytes; - size_t new_bytes = roundup_pow_of_two(new_top); - void *new_mem; + WARN_ON_ONCE(new_bytes > BTREE_TRANS_MEM_MAX); - WARN_ON_ONCE(new_bytes > BTREE_TRANS_MEM_MAX); + struct btree_transaction_stats *s = btree_trans_stats(trans); + if (s) + s->max_mem = max(s->max_mem, new_bytes); + + new_mem = krealloc(trans->mem, new_bytes, GFP_NOWAIT|__GFP_NOWARN); + if (unlikely(!new_mem)) { + bch2_trans_unlock(trans); - new_mem = krealloc(trans->mem, new_bytes, GFP_NOFS); + new_mem = krealloc(trans->mem, new_bytes, GFP_KERNEL); if (!new_mem && new_bytes <= BTREE_TRANS_MEM_MAX) { - new_mem = mempool_alloc(&trans->c->btree_trans_mem_pool, GFP_KERNEL); + new_mem = mempool_alloc(&c->btree_trans_mem_pool, GFP_KERNEL); new_bytes = BTREE_TRANS_MEM_MAX; kfree(trans->mem); } if (!new_mem) - return ERR_PTR(-ENOMEM); + return ERR_PTR(-BCH_ERR_ENOMEM_trans_kmalloc); trans->mem = new_mem; trans->mem_bytes = new_bytes; - if (old_bytes) { - trace_trans_restart_mem_realloced(trans->ip, _RET_IP_, new_bytes); - btree_trans_restart(trans); - return ERR_PTR(-EINTR); - } + ret = bch2_trans_relock(trans); + if (ret) + return ERR_PTR(ret); + } + + trans->mem = new_mem; + trans->mem_bytes = new_bytes; + + if (old_bytes) { + trace_and_count(c, trans_restart_mem_realloced, trans, _RET_IP_, new_bytes); + return ERR_PTR(btree_trans_restart(trans, BCH_ERR_transaction_restart_mem_realloced)); } p = trans->mem + trans->mem_top; @@ -2628,106 +2787,206 @@ void *bch2_trans_kmalloc(struct btree_trans *trans, size_t size) return p; } +static inline void check_srcu_held_too_long(struct btree_trans *trans) +{ + WARN(trans->srcu_held && time_after(jiffies, trans->srcu_lock_time + HZ * 10), + "btree trans held srcu lock (delaying memory reclaim) for %lu seconds", + (jiffies - trans->srcu_lock_time) / HZ); +} + +void bch2_trans_srcu_unlock(struct btree_trans *trans) +{ + if (trans->srcu_held) { + struct bch_fs *c = trans->c; + struct btree_path *path; + unsigned i; + + trans_for_each_path(trans, path, i) + if (path->cached && !btree_node_locked(path, 0)) + path->l[0].b = ERR_PTR(-BCH_ERR_no_btree_node_srcu_reset); + + check_srcu_held_too_long(trans); + srcu_read_unlock(&c->btree_trans_barrier, trans->srcu_idx); + trans->srcu_held = false; + } +} + +static void bch2_trans_srcu_lock(struct btree_trans *trans) +{ + if (!trans->srcu_held) { + trans->srcu_idx = srcu_read_lock(&trans->c->btree_trans_barrier); + trans->srcu_lock_time = jiffies; + trans->srcu_held = true; + } +} + /** * bch2_trans_begin() - reset a transaction after a interrupted attempt * @trans: transaction to reset * - * While iterating over nodes or updating nodes a attempt to lock a btree - * node may return EINTR when the trylock fails. When this occurs - * bch2_trans_begin() should be called and the transaction retried. + * Returns: current restart counter, to be used with trans_was_restarted() + * + * While iterating over nodes or updating nodes a attempt to lock a btree node + * may return BCH_ERR_transaction_restart when the trylock fails. When this + * occurs bch2_trans_begin() should be called and the transaction retried. */ -void bch2_trans_begin(struct btree_trans *trans) +u32 bch2_trans_begin(struct btree_trans *trans) { - struct btree_insert_entry *i; struct btree_path *path; + unsigned i; + u64 now; - trans_for_each_update(trans, i) - __btree_path_put(i->path, true); + bch2_trans_reset_updates(trans); - trans->extra_journal_res = 0; - trans->nr_updates = 0; + trans->restart_count++; trans->mem_top = 0; + trans->journal_entries = NULL; - trans->hooks = NULL; - trans->extra_journal_entries = NULL; - trans->extra_journal_entry_u64s = 0; - - if (trans->fs_usage_deltas) { - trans->fs_usage_deltas->used = 0; - memset(&trans->fs_usage_deltas->memset_start, 0, - (void *) &trans->fs_usage_deltas->memset_end - - (void *) &trans->fs_usage_deltas->memset_start); - } - - trans_for_each_path(trans, path) { + trans_for_each_path(trans, path, i) { path->should_be_locked = false; + /* + * If the transaction wasn't restarted, we're presuming to be + * doing something new: dont keep iterators excpt the ones that + * are in use - except for the subvolumes btree: + */ + if (!trans->restarted && path->btree_id != BTREE_ID_subvolumes) + path->preserve = false; + /* * XXX: we probably shouldn't be doing this if the transaction * was restarted, but currently we still overflow transaction * iterators if we do that */ if (!path->ref && !path->preserve) - __bch2_path_free(trans, path); - else if (!path->ref) + __bch2_path_free(trans, i); + else path->preserve = false; } - bch2_trans_cond_resched(trans); + now = local_clock(); + if (!trans->restarted && + (need_resched() || + now - trans->last_begin_time > BTREE_TRANS_MAX_LOCK_HOLD_TIME_NS)) { + drop_locks_do(trans, (cond_resched(), 0)); + now = local_clock(); + } + trans->last_begin_time = now; + + if (unlikely(trans->srcu_held && + time_after(jiffies, trans->srcu_lock_time + msecs_to_jiffies(10)))) + bch2_trans_srcu_unlock(trans); - if (trans->restarted) + trans->last_begin_ip = _RET_IP_; + if (trans->restarted) { bch2_btree_path_traverse_all(trans); + trans->notrace_relock_fail = false; + } - trans->restarted = false; + return trans->restart_count; } -static void bch2_trans_alloc_paths(struct btree_trans *trans, struct bch_fs *c) -{ - size_t paths_bytes = sizeof(struct btree_path) * BTREE_ITER_MAX; - size_t updates_bytes = sizeof(struct btree_insert_entry) * BTREE_ITER_MAX; - void *p = NULL; +const char *bch2_btree_transaction_fns[BCH_TRANSACTIONS_NR]; - BUG_ON(trans->used_mempool); +unsigned bch2_trans_get_fn_idx(const char *fn) +{ + unsigned i; -#ifdef __KERNEL__ - p = this_cpu_xchg(c->btree_paths_bufs->path , NULL); -#endif - if (!p) - p = mempool_alloc(&trans->c->btree_paths_pool, GFP_NOFS); + for (i = 0; i < ARRAY_SIZE(bch2_btree_transaction_fns); i++) + if (!bch2_btree_transaction_fns[i] || + bch2_btree_transaction_fns[i] == fn) { + bch2_btree_transaction_fns[i] = fn; + return i; + } - trans->paths = p; p += paths_bytes; - trans->updates = p; p += updates_bytes; + pr_warn_once("BCH_TRANSACTIONS_NR not big enough!"); + return i; } -void bch2_trans_init(struct btree_trans *trans, struct bch_fs *c, - unsigned expected_nr_iters, - size_t expected_mem_bytes) +struct btree_trans *__bch2_trans_get(struct bch_fs *c, unsigned fn_idx) __acquires(&c->btree_trans_barrier) { + struct btree_trans *trans; + + if (IS_ENABLED(__KERNEL__)) { + trans = this_cpu_xchg(c->btree_trans_bufs->trans, NULL); + if (trans) { + memset(trans, 0, offsetof(struct btree_trans, list)); + goto got_trans; + } + } + + trans = mempool_alloc(&c->btree_trans_pool, GFP_NOFS); memset(trans, 0, sizeof(*trans)); - trans->c = c; - trans->ip = _RET_IP_; + closure_init_stack(&trans->ref); + + seqmutex_lock(&c->btree_trans_lock); + if (IS_ENABLED(CONFIG_BCACHEFS_DEBUG)) { + struct btree_trans *pos; + pid_t pid = current->pid; - bch2_trans_alloc_paths(trans, c); + trans->locking_wait.task = current; - if (expected_mem_bytes) { - trans->mem_bytes = roundup_pow_of_two(expected_mem_bytes); - trans->mem = kmalloc(trans->mem_bytes, GFP_KERNEL|__GFP_NOFAIL); + list_for_each_entry(pos, &c->btree_trans_list, list) { + struct task_struct *pos_task = READ_ONCE(pos->locking_wait.task); + /* + * We'd much prefer to be stricter here and completely + * disallow multiple btree_trans in the same thread - + * but the data move path calls bch2_write when we + * already have a btree_trans initialized. + */ + BUG_ON(pos_task && + pid == pos_task->pid && + bch2_trans_locked(pos)); - if (!unlikely(trans->mem)) { - trans->mem = mempool_alloc(&c->btree_trans_mem_pool, GFP_KERNEL); - trans->mem_bytes = BTREE_TRANS_MEM_MAX; + if (pos_task && pid < pos_task->pid) { + list_add_tail(&trans->list, &pos->list); + goto list_add_done; + } } } + list_add_tail(&trans->list, &c->btree_trans_list); +list_add_done: + seqmutex_unlock(&c->btree_trans_lock); +got_trans: + trans->c = c; + trans->last_begin_time = local_clock(); + trans->fn_idx = fn_idx; + trans->locking_wait.task = current; + trans->journal_replay_not_finished = + unlikely(!test_bit(JOURNAL_REPLAY_DONE, &c->journal.flags)) && + atomic_inc_not_zero(&c->journal_keys.ref); + trans->nr_paths = ARRAY_SIZE(trans->_paths); + trans->paths_allocated = trans->_paths_allocated; + trans->sorted = trans->_sorted; + trans->paths = trans->_paths; + trans->updates = trans->_updates; - trans->srcu_idx = srcu_read_lock(&c->btree_trans_barrier); + *trans_paths_nr(trans->paths) = BTREE_ITER_MAX; -#ifdef CONFIG_BCACHEFS_DEBUG - trans->pid = current->pid; - mutex_lock(&c->btree_trans_lock); - list_add(&trans->list, &c->btree_trans_list); - mutex_unlock(&c->btree_trans_lock); -#endif + trans->paths_allocated[0] = 1; + + if (fn_idx < BCH_TRANSACTIONS_NR) { + trans->fn = bch2_btree_transaction_fns[fn_idx]; + + struct btree_transaction_stats *s = &c->btree_transaction_stats[fn_idx]; + + if (s->max_mem) { + unsigned expected_mem_bytes = roundup_pow_of_two(s->max_mem); + + trans->mem = kmalloc(expected_mem_bytes, GFP_KERNEL); + if (likely(trans->mem)) + trans->mem_bytes = expected_mem_bytes; + } + + trans->nr_paths_max = s->nr_max_paths; + trans->journal_entries_size = s->journal_entries_size; + } + + trans->srcu_idx = srcu_read_lock(&c->btree_trans_barrier); + trans->srcu_lock_time = jiffies; + trans->srcu_held = true; + return trans; } static void check_btree_paths_leaked(struct btree_trans *trans) @@ -2735,24 +2994,25 @@ static void check_btree_paths_leaked(struct btree_trans *trans) #ifdef CONFIG_BCACHEFS_DEBUG struct bch_fs *c = trans->c; struct btree_path *path; + unsigned i; - trans_for_each_path(trans, path) + trans_for_each_path(trans, path, i) if (path->ref) goto leaked; return; leaked: - bch_err(c, "btree paths leaked from %pS!", (void *) trans->ip); - trans_for_each_path(trans, path) + bch_err(c, "btree paths leaked from %s!", trans->fn); + trans_for_each_path(trans, path, i) if (path->ref) printk(KERN_ERR " btree %s %pS\n", - bch2_btree_ids[path->btree_id], + bch2_btree_id_str(path->btree_id), (void *) path->ip_allocated); /* Be noisy about this: */ bch2_fatal_error(c); #endif } -void bch2_trans_exit(struct btree_trans *trans) +void bch2_trans_put(struct btree_trans *trans) __releases(&c->btree_trans_barrier) { struct btree_insert_entry *i; @@ -2761,20 +3021,16 @@ void bch2_trans_exit(struct btree_trans *trans) bch2_trans_unlock(trans); trans_for_each_update(trans, i) - __btree_path_put(i->path, true); - trans->nr_updates = 0; + __btree_path_put(trans->paths + i->path, true); + trans->nr_updates = 0; + trans->locking_wait.task = NULL; check_btree_paths_leaked(trans); -#ifdef CONFIG_BCACHEFS_DEBUG - mutex_lock(&c->btree_trans_lock); - list_del(&trans->list); - mutex_unlock(&c->btree_trans_lock); -#endif - - srcu_read_unlock(&c->btree_trans_barrier, trans->srcu_idx); - - bch2_journal_preres_put(&c->journal, &trans->journal_preres); + if (trans->srcu_held) { + check_srcu_held_too_long(trans); + srcu_read_unlock(&c->btree_trans_barrier, trans->srcu_idx); + } if (trans->fs_usage_deltas) { if (trans->fs_usage_deltas->size + sizeof(trans->fs_usage_deltas) == @@ -2785,124 +3041,193 @@ void bch2_trans_exit(struct btree_trans *trans) kfree(trans->fs_usage_deltas); } + if (unlikely(trans->journal_replay_not_finished)) + bch2_journal_keys_put(c); + + unsigned long *paths_allocated = trans->paths_allocated; + trans->paths_allocated = NULL; + trans->paths = NULL; + + if (paths_allocated != trans->_paths_allocated) + kfree_rcu_mightsleep(paths_allocated); + if (trans->mem_bytes == BTREE_TRANS_MEM_MAX) mempool_free(trans->mem, &c->btree_trans_mem_pool); else kfree(trans->mem); -#ifdef __KERNEL__ - /* - * Userspace doesn't have a real percpu implementation: - */ - trans->paths = this_cpu_xchg(c->btree_paths_bufs->path, trans->paths); -#endif + /* Userspace doesn't have a real percpu implementation: */ + if (IS_ENABLED(__KERNEL__)) + trans = this_cpu_xchg(c->btree_trans_bufs->trans, trans); - if (trans->paths) - mempool_free(trans->paths, &c->btree_paths_pool); + if (trans) { + closure_sync(&trans->ref); - trans->mem = (void *) 0x1; - trans->paths = (void *) 0x1; + seqmutex_lock(&c->btree_trans_lock); + list_del(&trans->list); + seqmutex_unlock(&c->btree_trans_lock); + + mempool_free(trans, &c->btree_trans_pool); + } } static void __maybe_unused -bch2_btree_path_node_to_text(struct printbuf *out, - struct btree_bkey_cached_common *_b, - bool cached) +bch2_btree_bkey_cached_common_to_text(struct printbuf *out, + struct btree_bkey_cached_common *b) { - pr_buf(out, " l=%u %s:", - _b->level, bch2_btree_ids[_b->btree_id]); - bch2_bpos_to_text(out, btree_node_pos(_b, cached)); -} + struct six_lock_count c = six_lock_counts(&b->lock); + struct task_struct *owner; + pid_t pid; -#ifdef CONFIG_BCACHEFS_DEBUG -static bool trans_has_locks(struct btree_trans *trans) -{ - struct btree_path *path; + rcu_read_lock(); + owner = READ_ONCE(b->lock.owner); + pid = owner ? owner->pid : 0; + rcu_read_unlock(); - trans_for_each_path(trans, path) - if (path->nodes_locked) - return true; - return false; + prt_tab(out); + prt_printf(out, "%px %c l=%u %s:", b, b->cached ? 'c' : 'b', + b->level, bch2_btree_id_str(b->btree_id)); + bch2_bpos_to_text(out, btree_node_pos(b)); + + prt_tab(out); + prt_printf(out, " locks %u:%u:%u held by pid %u", + c.n[0], c.n[1], c.n[2], pid); } -#endif -void bch2_btree_trans_to_text(struct printbuf *out, struct bch_fs *c) +void bch2_btree_trans_to_text(struct printbuf *out, struct btree_trans *trans) { -#ifdef CONFIG_BCACHEFS_DEBUG - struct btree_trans *trans; - struct btree_path *path; - struct btree *b; - unsigned l; + struct btree_bkey_cached_common *b; + static char lock_types[] = { 'r', 'i', 'w' }; + struct task_struct *task = READ_ONCE(trans->locking_wait.task); + unsigned l, idx; - mutex_lock(&c->btree_trans_lock); - list_for_each_entry(trans, &c->btree_trans_list, list) { - if (!trans_has_locks(trans)) - continue; + /* before rcu_read_lock(): */ + bch2_printbuf_make_room(out, 4096); - pr_buf(out, "%i %ps\n", trans->pid, (void *) trans->ip); + if (!out->nr_tabstops) { + printbuf_tabstop_push(out, 16); + printbuf_tabstop_push(out, 32); + } - trans_for_each_path(trans, path) { - if (!path->nodes_locked) - continue; + prt_printf(out, "%i %s\n", task ? task->pid : 0, trans->fn); - pr_buf(out, " path %u %c l=%u %s:", - path->idx, - path->cached ? 'c' : 'b', - path->level, - bch2_btree_ids[path->btree_id]); - bch2_bpos_to_text(out, path->pos); - pr_buf(out, "\n"); - - for (l = 0; l < BTREE_MAX_DEPTH; l++) { - if (btree_node_locked(path, l)) { - pr_buf(out, " %s l=%u ", - btree_node_intent_locked(path, l) ? "i" : "r", l); - bch2_btree_path_node_to_text(out, - (void *) path->l[l].b, - path->cached); - pr_buf(out, "\n"); - } + /* trans->paths is rcu protected vs. freeing */ + rcu_read_lock(); + out->atomic++; + + struct btree_path *paths = rcu_dereference(trans->paths); + if (!paths) + goto out; + + unsigned long *paths_allocated = trans_paths_allocated(paths); + + trans_for_each_path_idx_from(paths_allocated, *trans_paths_nr(paths), idx, 1) { + struct btree_path *path = paths + idx; + if (!path->nodes_locked) + continue; + + prt_printf(out, " path %u %c l=%u %s:", + idx, + path->cached ? 'c' : 'b', + path->level, + bch2_btree_id_str(path->btree_id)); + bch2_bpos_to_text(out, path->pos); + prt_newline(out); + + for (l = 0; l < BTREE_MAX_DEPTH; l++) { + if (btree_node_locked(path, l) && + !IS_ERR_OR_NULL(b = (void *) READ_ONCE(path->l[l].b))) { + prt_printf(out, " %c l=%u ", + lock_types[btree_node_locked_type(path, l)], l); + bch2_btree_bkey_cached_common_to_text(out, b); + prt_newline(out); } } + } - b = READ_ONCE(trans->locking); - if (b) { - path = &trans->paths[trans->locking_path_idx]; - pr_buf(out, " locking path %u %c l=%u %s:", - trans->locking_path_idx, - path->cached ? 'c' : 'b', - trans->locking_level, - bch2_btree_ids[trans->locking_btree_id]); - bch2_bpos_to_text(out, trans->locking_pos); - - pr_buf(out, " node "); - bch2_btree_path_node_to_text(out, - (void *) b, path->cached); - pr_buf(out, "\n"); - } + b = READ_ONCE(trans->locking); + if (b) { + prt_printf(out, " blocked for %lluus on", + div_u64(local_clock() - trans->locking_wait.start_time, + 1000)); + prt_newline(out); + prt_printf(out, " %c", lock_types[trans->locking_wait.lock_want]); + bch2_btree_bkey_cached_common_to_text(out, b); + prt_newline(out); } - mutex_unlock(&c->btree_trans_lock); -#endif +out: + --out->atomic; + rcu_read_unlock(); } void bch2_fs_btree_iter_exit(struct bch_fs *c) { + struct btree_transaction_stats *s; + struct btree_trans *trans; + int cpu; + + if (c->btree_trans_bufs) + for_each_possible_cpu(cpu) { + struct btree_trans *trans = + per_cpu_ptr(c->btree_trans_bufs, cpu)->trans; + + if (trans) { + closure_sync(&trans->ref); + + seqmutex_lock(&c->btree_trans_lock); + list_del(&trans->list); + seqmutex_unlock(&c->btree_trans_lock); + } + kfree(trans); + } + free_percpu(c->btree_trans_bufs); + + trans = list_first_entry_or_null(&c->btree_trans_list, struct btree_trans, list); + if (trans) + panic("%s leaked btree_trans\n", trans->fn); + + for (s = c->btree_transaction_stats; + s < c->btree_transaction_stats + ARRAY_SIZE(c->btree_transaction_stats); + s++) { + kfree(s->max_paths_text); + bch2_time_stats_exit(&s->lock_hold_times); + } + + if (c->btree_trans_barrier_initialized) + cleanup_srcu_struct(&c->btree_trans_barrier); mempool_exit(&c->btree_trans_mem_pool); - mempool_exit(&c->btree_paths_pool); - cleanup_srcu_struct(&c->btree_trans_barrier); + mempool_exit(&c->btree_trans_pool); } -int bch2_fs_btree_iter_init(struct bch_fs *c) +void bch2_fs_btree_iter_init_early(struct bch_fs *c) { - unsigned nr = BTREE_ITER_MAX; + struct btree_transaction_stats *s; + + for (s = c->btree_transaction_stats; + s < c->btree_transaction_stats + ARRAY_SIZE(c->btree_transaction_stats); + s++) { + bch2_time_stats_init(&s->lock_hold_times); + mutex_init(&s->lock); + } INIT_LIST_HEAD(&c->btree_trans_list); - mutex_init(&c->btree_trans_lock); + seqmutex_init(&c->btree_trans_lock); +} + +int bch2_fs_btree_iter_init(struct bch_fs *c) +{ + int ret; + + c->btree_trans_bufs = alloc_percpu(struct btree_trans_buf); + if (!c->btree_trans_bufs) + return -ENOMEM; - return init_srcu_struct(&c->btree_trans_barrier) ?: - mempool_init_kmalloc_pool(&c->btree_paths_pool, 1, - sizeof(struct btree_path) * nr + - sizeof(struct btree_insert_entry) * nr) ?: + ret = mempool_init_kmalloc_pool(&c->btree_trans_pool, 1, + sizeof(struct btree_trans)) ?: mempool_init_kmalloc_pool(&c->btree_trans_mem_pool, 1, - BTREE_TRANS_MEM_MAX); + BTREE_TRANS_MEM_MAX) ?: + init_srcu_struct(&c->btree_trans_barrier); + if (!ret) + c->btree_trans_barrier_initialized = true; + return ret; }