X-Git-Url: https://git.sesse.net/?a=blobdiff_plain;f=libbcachefs%2Fbtree_iter.c;h=c356032be3ec43938eed81339740abde9d247e13;hb=55142cd0b5ef2a2150d4708dad0c3fd54a3ffd39;hp=33cbc2ff5c9a1ebc99934c0ee8ff7a4594f7224c;hpb=d13bbb2955f899f10be4ab315ad229d2951fdcda;p=bcachefs-tools-debian diff --git a/libbcachefs/btree_iter.c b/libbcachefs/btree_iter.c index 33cbc2f..c356032 100644 --- a/libbcachefs/btree_iter.c +++ b/libbcachefs/btree_iter.c @@ -1,78 +1,98 @@ +// SPDX-License-Identifier: GPL-2.0 #include "bcachefs.h" #include "bkey_methods.h" +#include "bkey_buf.h" #include "btree_cache.h" #include "btree_iter.h" +#include "btree_key_cache.h" #include "btree_locking.h" +#include "btree_update.h" #include "debug.h" +#include "error.h" #include "extents.h" +#include "journal.h" +#include "replicas.h" #include #include -static inline struct bkey_s_c __btree_iter_peek_all(struct btree_iter *, - struct btree_iter_level *, - struct bkey *); +static void btree_iter_set_search_pos(struct btree_iter *, struct bpos); +static struct btree_iter *btree_iter_child_alloc(struct btree_iter *, unsigned long); +static struct btree_iter *btree_trans_iter_alloc(struct btree_trans *); +static void btree_iter_copy(struct btree_iter *, struct btree_iter *); -#define BTREE_ITER_NOT_END ((struct btree *) 1) +static inline struct bpos bkey_successor(struct btree_iter *iter, struct bpos p) +{ + EBUG_ON(btree_iter_type(iter) == BTREE_ITER_NODES); + + /* Are we iterating over keys in all snapshots? */ + if (iter->flags & BTREE_ITER_ALL_SNAPSHOTS) { + p = bpos_successor(p); + } else { + p = bpos_nosnap_successor(p); + p.snapshot = iter->snapshot; + } + + return p; +} + +static inline struct bpos bkey_predecessor(struct btree_iter *iter, struct bpos p) +{ + EBUG_ON(btree_iter_type(iter) == BTREE_ITER_NODES); + + /* Are we iterating over keys in all snapshots? */ + if (iter->flags & BTREE_ITER_ALL_SNAPSHOTS) { + p = bpos_predecessor(p); + } else { + p = bpos_nosnap_predecessor(p); + p.snapshot = iter->snapshot; + } + + return p; +} static inline bool is_btree_node(struct btree_iter *iter, unsigned l) { return l < BTREE_MAX_DEPTH && - iter->l[l].b && - iter->l[l].b != BTREE_ITER_NOT_END; + (unsigned long) iter->l[l].b >= 128; } -/* Returns < 0 if @k is before iter pos, > 0 if @k is after */ -static inline int __btree_iter_pos_cmp(struct btree_iter *iter, - const struct btree *b, - const struct bkey_packed *k, - bool interior_node) +static inline struct bpos btree_iter_search_key(struct btree_iter *iter) { - int cmp = bkey_cmp_left_packed(b, k, &iter->pos); + struct bpos pos = iter->pos; - if (cmp) - return cmp; - if (bkey_deleted(k)) - return -1; + if ((iter->flags & BTREE_ITER_IS_EXTENTS) && + bkey_cmp(pos, POS_MAX)) + pos = bkey_successor(iter, pos); + return pos; +} - /* - * Normally, for extents we want the first key strictly greater than - * the iterator position - with the exception that for interior nodes, - * we don't want to advance past the last key if the iterator position - * is POS_MAX: - */ - if (iter->flags & BTREE_ITER_IS_EXTENTS && - (!interior_node || - bkey_cmp_left_packed_byval(b, k, POS_MAX))) - return -1; - return 1; +static inline bool btree_iter_pos_before_node(struct btree_iter *iter, + struct btree *b) +{ + return bpos_cmp(iter->real_pos, b->data->min_key) < 0; +} + +static inline bool btree_iter_pos_after_node(struct btree_iter *iter, + struct btree *b) +{ + return bpos_cmp(b->key.k.p, iter->real_pos) < 0; } -static inline int btree_iter_pos_cmp(struct btree_iter *iter, - const struct btree *b, - const struct bkey_packed *k) +static inline bool btree_iter_pos_in_node(struct btree_iter *iter, + struct btree *b) { - return __btree_iter_pos_cmp(iter, b, k, b->level != 0); + return iter->btree_id == b->c.btree_id && + !btree_iter_pos_before_node(iter, b) && + !btree_iter_pos_after_node(iter, b); } /* Btree node locking: */ -/* - * Updates the saved lock sequence number, so that bch2_btree_node_relock() will - * succeed: - */ void bch2_btree_node_unlock_write(struct btree *b, struct btree_iter *iter) { - struct btree_iter *linked; - - EBUG_ON(iter->l[b->level].b != b); - EBUG_ON(iter->l[b->level].lock_seq + 1 != b->lock.state.seq); - - trans_for_each_iter_with_node(iter->trans, b, linked) - linked->l[b->level].lock_seq += 2; - - six_unlock_write(&b->lock); + bch2_btree_node_unlock_write_inlined(b, iter); } void __bch2_btree_node_lock_write(struct btree *b, struct btree_iter *iter) @@ -80,11 +100,11 @@ void __bch2_btree_node_lock_write(struct btree *b, struct btree_iter *iter) struct btree_iter *linked; unsigned readers = 0; - EBUG_ON(btree_node_read_locked(iter, b->level)); + EBUG_ON(!btree_node_intent_locked(iter, b->c.level)); trans_for_each_iter(iter->trans, linked) - if (linked->l[b->level].b == b && - btree_node_read_locked(linked, b->level)) + if (linked->l[b->c.level].b == b && + btree_node_read_locked(linked, b->c.level)) readers++; /* @@ -94,10 +114,10 @@ void __bch2_btree_node_lock_write(struct btree *b, struct btree_iter *iter) * locked: */ atomic64_sub(__SIX_VAL(read_lock, readers), - &b->lock.state.counter); + &b->c.lock.state.counter); btree_node_lock_type(iter->trans->c, b, SIX_LOCK_write); atomic64_add(__SIX_VAL(read_lock, readers), - &b->lock.state.counter); + &b->c.lock.state.counter); } bool __bch2_btree_node_relock(struct btree_iter *iter, unsigned level) @@ -105,19 +125,20 @@ bool __bch2_btree_node_relock(struct btree_iter *iter, unsigned level) struct btree *b = btree_iter_node(iter, level); int want = __btree_lock_want(iter, level); - if (!b || b == BTREE_ITER_NOT_END) + if (!is_btree_node(iter, level)) return false; if (race_fault()) return false; - if (!six_relock_type(&b->lock, want, iter->l[level].lock_seq) && - !(iter->l[level].lock_seq >> 1 == b->lock.state.seq >> 1 && - btree_node_lock_increment(iter, b, level, want))) + if (six_relock_type(&b->c.lock, want, iter->l[level].lock_seq) || + (btree_node_lock_seq_matches(iter, b, level) && + btree_node_lock_increment(iter->trans, b, level, want))) { + mark_btree_node_locked(iter, level, want); + return true; + } else { return false; - - mark_btree_node_locked(iter, level, want); - return true; + } } static bool bch2_btree_node_upgrade(struct btree_iter *iter, unsigned level) @@ -136,12 +157,12 @@ static bool bch2_btree_node_upgrade(struct btree_iter *iter, unsigned level) return false; if (btree_node_locked(iter, level) - ? six_lock_tryupgrade(&b->lock) - : six_relock_type(&b->lock, SIX_LOCK_intent, iter->l[level].lock_seq)) + ? six_lock_tryupgrade(&b->c.lock) + : six_relock_type(&b->c.lock, SIX_LOCK_intent, iter->l[level].lock_seq)) goto success; - if (iter->l[level].lock_seq >> 1 == b->lock.state.seq >> 1 && - btree_node_lock_increment(iter, b, level, BTREE_NODE_INTENT_LOCKED)) { + if (btree_node_lock_seq_matches(iter, b, level) && + btree_node_lock_increment(iter->trans, b, level, BTREE_NODE_INTENT_LOCKED)) { btree_node_unlock(iter, level); goto success; } @@ -152,8 +173,8 @@ success: return true; } -static inline bool btree_iter_get_locks(struct btree_iter *iter, - bool upgrade) +static inline bool btree_iter_get_locks(struct btree_iter *iter, bool upgrade, + unsigned long trace_ip) { unsigned l = iter->level; int fail_idx = -1; @@ -165,6 +186,18 @@ static inline bool btree_iter_get_locks(struct btree_iter *iter, if (!(upgrade ? bch2_btree_node_upgrade(iter, l) : bch2_btree_node_relock(iter, l))) { + (upgrade + ? trace_node_upgrade_fail + : trace_node_relock_fail)(iter->trans->ip, trace_ip, + iter->btree_id, &iter->real_pos, + l, iter->l[l].lock_seq, + is_btree_node(iter, l) + ? 0 + : (unsigned long) iter->l[l].b, + is_btree_node(iter, l) + ? iter->l[l].b->c.lock.state.seq + : 0); + fail_idx = l; btree_iter_set_dirty(iter, BTREE_ITER_NEED_TRAVERSE); } @@ -179,7 +212,7 @@ static inline bool btree_iter_get_locks(struct btree_iter *iter, */ while (fail_idx >= 0) { btree_node_unlock(iter, fail_idx); - iter->l[fail_idx].b = BTREE_ITER_NOT_END; + iter->l[fail_idx].b = BTREE_ITER_NO_NODE_GET_LOCKS; --fail_idx; } @@ -191,25 +224,32 @@ static inline bool btree_iter_get_locks(struct btree_iter *iter, return iter->uptodate < BTREE_ITER_NEED_RELOCK; } +static struct bpos btree_node_pos(struct btree_bkey_cached_common *_b, + enum btree_iter_type type) +{ + return type != BTREE_ITER_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 *b, struct bpos pos, - unsigned level, - struct btree_iter *iter, - enum six_lock_type type, - bool may_drop_locks) + unsigned level, struct btree_iter *iter, + enum six_lock_type type, + six_lock_should_sleep_fn should_sleep_fn, void *p, + unsigned long ip) { - struct btree_iter *linked; - bool ret = true; + struct btree_trans *trans = iter->trans; + struct btree_iter *linked, *deadlock_iter = NULL; + u64 start_time = local_clock(); + unsigned reason = 9; + bool ret; /* Check if it's safe to block: */ - trans_for_each_iter(iter->trans, linked) { + trans_for_each_iter(trans, linked) { if (!linked->nodes_locked) continue; - /* * Must lock btree nodes in key order: */ - if (__btree_iter_cmp(iter->btree_id, pos, linked) < 0) - ret = false; - /* * Can't block taking an intent lock if we have _any_ nodes read * locked: @@ -224,13 +264,28 @@ bool __bch2_btree_node_lock(struct btree *b, struct bpos pos, */ if (type == SIX_LOCK_intent && linked->nodes_locked != linked->nodes_intent_locked) { - if (may_drop_locks) { - linked->locks_want = max_t(unsigned, - linked->locks_want, - __fls(linked->nodes_locked) + 1); - btree_iter_get_locks(linked, true); + deadlock_iter = linked; + reason = 1; + } + + if (linked->btree_id != iter->btree_id) { + if (linked->btree_id > iter->btree_id) { + deadlock_iter = linked; + reason = 3; + } + continue; + } + + /* + * Within the same btree, cached iterators come before non + * cached iterators: + */ + if (btree_iter_is_cached(linked) != btree_iter_is_cached(iter)) { + if (btree_iter_is_cached(iter)) { + deadlock_iter = linked; + reason = 4; } - ret = false; + continue; } /* @@ -238,38 +293,68 @@ bool __bch2_btree_node_lock(struct btree *b, struct bpos pos, * another iterator has possible descendants locked of the node * we're about to lock, it must have the ancestors locked too: */ - if (linked->btree_id == iter->btree_id && - level > __fls(linked->nodes_locked)) { - if (may_drop_locks) { - linked->locks_want = - max(level + 1, max_t(unsigned, - linked->locks_want, - iter->locks_want)); - btree_iter_get_locks(linked, true); - } - ret = false; + if (level > __fls(linked->nodes_locked)) { + deadlock_iter = 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, + btree_iter_type(linked))) <= 0) { + deadlock_iter = linked; + reason = 7; + BUG_ON(trans->in_traverse_all); } } - if (ret) - __btree_node_lock_type(iter->trans->c, b, type); - else - trans_restart(); + if (unlikely(deadlock_iter)) { + trace_trans_restart_would_deadlock(iter->trans->ip, ip, + trans->in_traverse_all, reason, + deadlock_iter->btree_id, + btree_iter_type(deadlock_iter), + &deadlock_iter->real_pos, + iter->btree_id, + btree_iter_type(iter), + &pos); + return false; + } + + if (six_trylock_type(&b->c.lock, type)) + return true; + +#ifdef CONFIG_BCACHEFS_DEBUG + trans->locking_iter_idx = iter->idx; + trans->locking_pos = pos; + trans->locking_btree_id = iter->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 -void bch2_btree_iter_verify_locks(struct btree_iter *iter) +static void bch2_btree_iter_verify_locks(struct btree_iter *iter) { unsigned l; - BUG_ON((iter->flags & BTREE_ITER_NOUNLOCK) && - !btree_node_locked(iter, 0)); + if (!(iter->trans->iters_linked & (1ULL << iter->idx))) { + BUG_ON(iter->nodes_locked); + return; + } - for (l = 0; btree_iter_node(iter, l); l++) { + for (l = 0; is_btree_node(iter, l); l++) { if (iter->uptodate >= BTREE_ITER_NEED_RELOCK && !btree_node_locked(iter, l)) continue; @@ -286,14 +371,14 @@ void bch2_btree_trans_verify_locks(struct btree_trans *trans) trans_for_each_iter(trans, iter) bch2_btree_iter_verify_locks(iter); } +#else +static inline void bch2_btree_iter_verify_locks(struct btree_iter *iter) {} #endif __flatten -static bool bch2_btree_iter_relock(struct btree_iter *iter) +static bool bch2_btree_iter_relock(struct btree_iter *iter, unsigned long trace_ip) { - return iter->uptodate >= BTREE_ITER_NEED_RELOCK - ? btree_iter_get_locks(iter, false) - : true; + return btree_iter_get_locks(iter, false, trace_ip); } bool __bch2_btree_iter_upgrade(struct btree_iter *iter, @@ -305,109 +390,97 @@ bool __bch2_btree_iter_upgrade(struct btree_iter *iter, iter->locks_want = new_locks_want; - if (btree_iter_get_locks(iter, true)) + if (btree_iter_get_locks(iter, true, _THIS_IP_)) return true; /* - * Ancestor nodes must be locked before child nodes, so set locks_want - * on iterators that might lock ancestors before us to avoid getting - * -EINTR later: + * 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_iter_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_iter_traverse_all(). */ trans_for_each_iter(iter->trans, linked) if (linked != iter && + btree_iter_type(linked) == btree_iter_type(iter) && linked->btree_id == iter->btree_id && - btree_iter_cmp(linked, iter) <= 0 && linked->locks_want < new_locks_want) { linked->locks_want = new_locks_want; - btree_iter_get_locks(linked, true); + btree_iter_get_locks(linked, true, _THIS_IP_); } return false; } -bool __bch2_btree_iter_upgrade_nounlock(struct btree_iter *iter, - unsigned new_locks_want) -{ - unsigned l = iter->level; - - EBUG_ON(iter->locks_want >= new_locks_want); - - iter->locks_want = new_locks_want; - - do { - if (!btree_iter_node(iter, l)) - break; - - if (!bch2_btree_node_upgrade(iter, l)) { - iter->locks_want = l; - return false; - } - - l++; - } while (l < iter->locks_want); - - return true; -} - void __bch2_btree_iter_downgrade(struct btree_iter *iter, - unsigned downgrade_to) + unsigned new_locks_want) { - struct btree_iter *linked; unsigned l; - /* - * We downgrade linked iterators as well because btree_iter_upgrade - * might have had to modify locks_want on linked iterators due to lock - * ordering: - */ - trans_for_each_iter(iter->trans, linked) { - unsigned new_locks_want = downgrade_to ?: - (linked->flags & BTREE_ITER_INTENT ? 1 : 0); - - if (linked->locks_want <= new_locks_want) - continue; + EBUG_ON(iter->locks_want < new_locks_want); - linked->locks_want = new_locks_want; + iter->locks_want = new_locks_want; - while (linked->nodes_locked && - (l = __fls(linked->nodes_locked)) >= linked->locks_want) { - if (l > linked->level) { - btree_node_unlock(linked, l); - } else { - if (btree_node_intent_locked(linked, l)) { - six_lock_downgrade(&linked->l[l].b->lock); - linked->nodes_intent_locked ^= 1 << l; - } - break; + while (iter->nodes_locked && + (l = __fls(iter->nodes_locked)) >= iter->locks_want) { + if (l > iter->level) { + btree_node_unlock(iter, l); + } else { + if (btree_node_intent_locked(iter, l)) { + six_lock_downgrade(&iter->l[l].b->c.lock); + iter->nodes_intent_locked ^= 1 << l; } + break; } } bch2_btree_trans_verify_locks(iter->trans); } -int bch2_btree_iter_unlock(struct btree_iter *iter) +void bch2_trans_downgrade(struct btree_trans *trans) { - struct btree_iter *linked; + struct btree_iter *iter; - trans_for_each_iter(iter->trans, linked) - __bch2_btree_iter_unlock(linked); + trans_for_each_iter(trans, iter) + bch2_btree_iter_downgrade(iter); +} - return btree_iter_err(iter); +/* Btree transaction locking: */ + +static inline bool btree_iter_should_be_locked(struct btree_trans *trans, + struct btree_iter *iter) +{ + return (iter->flags & BTREE_ITER_KEEP_UNTIL_COMMIT) || + iter->should_be_locked; } -bool bch2_btree_trans_relock(struct btree_trans *trans) +bool bch2_trans_relock(struct btree_trans *trans) { struct btree_iter *iter; - bool ret = true; trans_for_each_iter(trans, iter) - ret &= bch2_btree_iter_relock(iter); - - return ret; + if (!bch2_btree_iter_relock(iter, _RET_IP_) && + btree_iter_should_be_locked(trans, iter)) { + trace_trans_restart_relock(trans->ip, _RET_IP_, + iter->btree_id, &iter->real_pos); + return false; + } + return true; } -void bch2_btree_trans_unlock(struct btree_trans *trans) +void bch2_trans_unlock(struct btree_trans *trans) { struct btree_iter *iter; @@ -415,26 +488,67 @@ void bch2_btree_trans_unlock(struct btree_trans *trans) __bch2_btree_iter_unlock(iter); } -/* Btree transaction locking: */ - /* Btree iterator: */ #ifdef CONFIG_BCACHEFS_DEBUG -static void __bch2_btree_iter_verify(struct btree_iter *iter, - struct btree *b) +static void bch2_btree_iter_verify_cached(struct btree_iter *iter) { - struct btree_iter_level *l = &iter->l[b->level]; - struct btree_node_iter tmp = l->iter; - struct bkey_packed *k; + struct bkey_cached *ck; + bool locked = btree_node_locked(iter, 0); + + if (!bch2_btree_node_relock(iter, 0)) + return; + + ck = (void *) iter->l[0].b; + BUG_ON(ck->key.btree_id != iter->btree_id || + bkey_cmp(ck->key.pos, iter->pos)); + + if (!locked) + btree_node_unlock(iter, 0); +} - if (!debug_check_iterators(iter->trans->c)) +static void bch2_btree_iter_verify_level(struct btree_iter *iter, + unsigned level) +{ + struct btree_iter_level *l; + struct btree_node_iter tmp; + bool locked; + struct bkey_packed *p, *k; + char buf1[100], buf2[100], buf3[100]; + const char *msg; + + if (!bch2_debug_check_iterators) + return; + + l = &iter->l[level]; + tmp = l->iter; + locked = btree_node_locked(iter, level); + + if (btree_iter_type(iter) == BTREE_ITER_CACHED) { + if (!level) + bch2_btree_iter_verify_cached(iter); + return; + } + + BUG_ON(iter->level < iter->min_depth); + + if (!btree_iter_node(iter, level)) return; - if (iter->uptodate > BTREE_ITER_NEED_PEEK) + if (!bch2_btree_node_relock(iter, level)) return; - bch2_btree_node_iter_verify(&l->iter, b); + BUG_ON(!btree_iter_pos_in_node(iter, l->b)); + + /* + * node iterators don't use leaf node iterator: + */ + if (btree_iter_type(iter) == BTREE_ITER_NODES && + level <= iter->min_depth) + goto unlock; + + bch2_btree_node_iter_verify(&l->iter, l->b); /* * For interior nodes, the iterator will have skipped past @@ -443,54 +557,149 @@ static void __bch2_btree_iter_verify(struct btree_iter *iter, * For extents, the iterator may have skipped past deleted keys (but not * whiteouts) */ - k = b->level || iter->flags & BTREE_ITER_IS_EXTENTS - ? bch2_btree_node_iter_prev_filter(&tmp, b, KEY_TYPE_discard) - : bch2_btree_node_iter_prev_all(&tmp, b); - if (k && btree_iter_pos_cmp(iter, b, k) > 0) { - char buf[100]; - struct bkey uk = bkey_unpack_key(b, k); + p = level || btree_node_type_is_extents(iter->btree_id) + ? bch2_btree_node_iter_prev(&tmp, l->b) + : bch2_btree_node_iter_prev_all(&tmp, l->b); + k = bch2_btree_node_iter_peek_all(&l->iter, l->b); - bch2_bkey_to_text(&PBUF(buf), &uk); - panic("prev key should be before iter pos:\n%s\n%llu:%llu\n", - buf, iter->pos.inode, iter->pos.offset); + if (p && bkey_iter_pos_cmp(l->b, p, &iter->real_pos) >= 0) { + msg = "before"; + goto err; } - k = bch2_btree_node_iter_peek_all(&l->iter, b); - if (k && btree_iter_pos_cmp(iter, b, k) < 0) { - char buf[100]; - struct bkey uk = bkey_unpack_key(b, k); + if (k && bkey_iter_pos_cmp(l->b, k, &iter->real_pos) < 0) { + msg = "after"; + goto err; + } +unlock: + if (!locked) + btree_node_unlock(iter, level); + return; +err: + strcpy(buf2, "(none)"); + strcpy(buf3, "(none)"); - bch2_bkey_to_text(&PBUF(buf), &uk); - panic("iter should be after current key:\n" - "iter pos %llu:%llu\n" - "cur key %s\n", - iter->pos.inode, iter->pos.offset, buf); + bch2_bpos_to_text(&PBUF(buf1), iter->real_pos); + + if (p) { + struct bkey uk = bkey_unpack_key(l->b, p); + bch2_bkey_to_text(&PBUF(buf2), &uk); + } + + if (k) { + struct bkey uk = bkey_unpack_key(l->b, k); + bch2_bkey_to_text(&PBUF(buf3), &uk); } - BUG_ON(iter->uptodate == BTREE_ITER_UPTODATE && - (iter->flags & BTREE_ITER_TYPE) == BTREE_ITER_KEYS && - !bkey_whiteout(&iter->k) && - bch2_btree_node_iter_end(&l->iter)); + panic("iterator should be %s key at level %u:\n" + "iter pos %s\n" + "prev key %s\n" + "cur key %s\n", + msg, level, buf1, buf2, buf3); } -void bch2_btree_iter_verify(struct btree_iter *iter, struct btree *b) +static void bch2_btree_iter_verify(struct btree_iter *iter) { - struct btree_iter *linked; + enum btree_iter_type type = btree_iter_type(iter); + unsigned i; + + EBUG_ON(iter->btree_id >= BTREE_ID_NR); + + BUG_ON(!(iter->flags & BTREE_ITER_ALL_SNAPSHOTS) && + iter->pos.snapshot != iter->snapshot); + + BUG_ON((iter->flags & BTREE_ITER_IS_EXTENTS) && + (iter->flags & BTREE_ITER_ALL_SNAPSHOTS)); + + BUG_ON(type == BTREE_ITER_NODES && + !(iter->flags & BTREE_ITER_ALL_SNAPSHOTS)); + + BUG_ON(type != BTREE_ITER_NODES && + (iter->flags & BTREE_ITER_ALL_SNAPSHOTS) && + !btree_type_has_snapshots(iter->btree_id)); + + bch2_btree_iter_verify_locks(iter); - if (!debug_check_iterators(iter->trans->c)) + for (i = 0; i < BTREE_MAX_DEPTH; i++) + bch2_btree_iter_verify_level(iter, i); +} + +static void bch2_btree_iter_verify_entry_exit(struct btree_iter *iter) +{ + enum btree_iter_type type = btree_iter_type(iter); + + BUG_ON(!(iter->flags & BTREE_ITER_ALL_SNAPSHOTS) && + iter->pos.snapshot != iter->snapshot); + + BUG_ON((type == BTREE_ITER_KEYS || + type == BTREE_ITER_CACHED) && + (bkey_cmp(iter->pos, bkey_start_pos(&iter->k)) < 0 || + bkey_cmp(iter->pos, iter->k.p) > 0)); +} + +void bch2_btree_trans_verify_iters(struct btree_trans *trans, struct btree *b) +{ + struct btree_iter *iter; + + if (!bch2_debug_check_iterators) return; - trans_for_each_iter_with_node(iter->trans, b, linked) - __bch2_btree_iter_verify(linked, b); + trans_for_each_iter_with_node(trans, b, iter) + bch2_btree_iter_verify_level(iter, b->c.level); } #else -static inline void __bch2_btree_iter_verify(struct btree_iter *iter, - struct btree *b) {} +static inline void bch2_btree_iter_verify_level(struct btree_iter *iter, unsigned l) {} +static inline void bch2_btree_iter_verify(struct btree_iter *iter) {} +static inline void bch2_btree_iter_verify_entry_exit(struct btree_iter *iter) {} #endif +static void btree_node_iter_set_set_pos(struct btree_node_iter *iter, + struct btree *b, + struct bset_tree *t, + struct bkey_packed *k) +{ + struct btree_node_iter_set *set; + + btree_node_iter_for_each(iter, set) + if (set->end == t->end_offset) { + set->k = __btree_node_key_to_offset(b, k); + bch2_btree_node_iter_sort(iter, b); + return; + } + + bch2_btree_node_iter_push(iter, b, k, btree_bkey_last(b, t)); +} + +static void __bch2_btree_iter_fix_key_modified(struct btree_iter *iter, + struct btree *b, + struct bkey_packed *where) +{ + struct btree_iter_level *l = &iter->l[b->c.level]; + + if (where != bch2_btree_node_iter_peek_all(&l->iter, l->b)) + return; + + if (bkey_iter_pos_cmp(l->b, where, &iter->real_pos) < 0) + bch2_btree_node_iter_advance(&l->iter, l->b); + + btree_iter_set_dirty(iter, BTREE_ITER_NEED_PEEK); +} + +void bch2_btree_iter_fix_key_modified(struct btree_iter *iter, + struct btree *b, + struct bkey_packed *where) +{ + struct btree_iter *linked; + + trans_for_each_iter_with_node(iter->trans, b, linked) { + __bch2_btree_iter_fix_key_modified(linked, b, where); + bch2_btree_iter_verify_level(linked, b->c.level); + } +} + static void __bch2_btree_node_iter_fix(struct btree_iter *iter, struct btree *b, struct btree_node_iter *node_iter, @@ -504,6 +713,10 @@ static void __bch2_btree_node_iter_fix(struct btree_iter *iter, unsigned offset = __btree_node_key_to_offset(b, where); int shift = new_u64s - clobber_u64s; unsigned old_end = t->end_offset - shift; + unsigned orig_iter_pos = node_iter->data[0].k; + bool iter_current_key_modified = + orig_iter_pos >= offset && + orig_iter_pos <= offset + clobber_u64s; btree_node_iter_for_each(node_iter, set) if (set->end == old_end) @@ -511,18 +724,13 @@ static void __bch2_btree_node_iter_fix(struct btree_iter *iter, /* didn't find the bset in the iterator - might have to readd it: */ if (new_u64s && - btree_iter_pos_cmp(iter, b, where) > 0) { - btree_iter_set_dirty(iter, BTREE_ITER_NEED_PEEK); - + bkey_iter_pos_cmp(b, where, &iter->real_pos) >= 0) { bch2_btree_node_iter_push(node_iter, b, where, end); - - if (!b->level && - node_iter == &iter->l[0].iter) - bkey_disassemble(b, - bch2_btree_node_iter_peek_all(node_iter, b), - &iter->k); + goto fixup_done; + } else { + /* Iterator is after key that changed */ + return; } - return; found: set->end = t->end_offset; @@ -531,93 +739,63 @@ found: return; if (new_u64s && - btree_iter_pos_cmp(iter, b, where) > 0) { + bkey_iter_pos_cmp(b, where, &iter->real_pos) >= 0) { set->k = offset; } else if (set->k < offset + clobber_u64s) { set->k = offset + new_u64s; if (set->k == set->end) bch2_btree_node_iter_set_drop(node_iter, set); } else { + /* Iterator is after key that changed */ set->k = (int) set->k + shift; - goto iter_current_key_not_modified; + return; } - btree_iter_set_dirty(iter, BTREE_ITER_NEED_PEEK); - bch2_btree_node_iter_sort(node_iter, b); - if (!b->level && node_iter == &iter->l[0].iter) { - /* - * not legal to call bkey_debugcheck() here, because we're - * called midway through the update path after update has been - * marked but before deletes have actually happened: - */ -#if 0 - __btree_iter_peek_all(iter, &iter->l[0], &iter->k); -#endif - struct btree_iter_level *l = &iter->l[0]; - struct bkey_packed *k = - bch2_btree_node_iter_peek_all(&l->iter, l->b); - - if (unlikely(!k)) - iter->k.type = KEY_TYPE_deleted; - else - bkey_disassemble(l->b, k, &iter->k); - } -iter_current_key_not_modified: +fixup_done: + if (node_iter->data[0].k != orig_iter_pos) + iter_current_key_modified = true; /* - * Interior nodes are special because iterators for interior nodes don't - * obey the usual invariants regarding the iterator position: - * - * We may have whiteouts that compare greater than the iterator - * position, and logically should be in the iterator, but that we - * skipped past to find the first live key greater than the iterator - * position. This becomes an issue when we insert a new key that is - * greater than the current iterator position, but smaller than the - * whiteouts we've already skipped past - this happens in the course of - * a btree split. - * - * We have to rewind the iterator past to before those whiteouts here, - * else bkey_node_iter_prev() is not going to work and who knows what - * else would happen. And we have to do it manually, because here we've - * already done the insert and the iterator is currently inconsistent: - * - * We've got multiple competing invariants, here - we have to be careful - * about rewinding iterators for interior nodes, because they should - * always point to the key for the child node the btree iterator points - * to. + * When a new key is added, and the node iterator now points to that + * key, the iterator might have skipped past deleted keys that should + * come after the key the iterator now points to. We have to rewind to + * before those deleted keys - otherwise + * bch2_btree_node_iter_prev_all() breaks: */ - if (b->level && new_u64s && - btree_iter_pos_cmp(iter, b, where) > 0) { + if (!bch2_btree_node_iter_end(node_iter) && + iter_current_key_modified && + (b->c.level || + btree_node_type_is_extents(iter->btree_id))) { struct bset_tree *t; - struct bkey_packed *k; + struct bkey_packed *k, *k2, *p; + + k = bch2_btree_node_iter_peek_all(node_iter, b); for_each_bset(b, t) { - if (bch2_bkey_to_bset(b, where) == t) + bool set_pos = false; + + if (node_iter->data[0].end == t->end_offset) continue; - k = bch2_bkey_prev_all(b, t, - bch2_btree_node_iter_bset_pos(node_iter, b, t)); - if (k && - bkey_iter_cmp(b, k, where) > 0) { - struct btree_node_iter_set *set; - unsigned offset = - __btree_node_key_to_offset(b, bkey_next(k)); - - btree_node_iter_for_each(node_iter, set) - if (set->k == offset) { - set->k = __btree_node_key_to_offset(b, k); - bch2_btree_node_iter_sort(node_iter, b); - goto next_bset; - } - - bch2_btree_node_iter_push(node_iter, b, k, - btree_bkey_last(b, t)); + k2 = bch2_btree_node_iter_bset_pos(node_iter, b, t); + + while ((p = bch2_bkey_prev_all(b, t, k2)) && + bkey_iter_cmp(b, k, p) < 0) { + k2 = p; + set_pos = true; } -next_bset: - t = t; + + if (set_pos) + btree_node_iter_set_set_pos(node_iter, + b, t, k2); } } + + if (!b->c.level && + node_iter == &iter->l[0].iter && + iter_current_key_modified) + btree_iter_set_dirty(iter, BTREE_ITER_NEED_PEEK); } void bch2_btree_node_iter_fix(struct btree_iter *iter, @@ -630,14 +808,20 @@ void bch2_btree_node_iter_fix(struct btree_iter *iter, struct bset_tree *t = bch2_bkey_to_bset(b, where); struct btree_iter *linked; - if (node_iter != &iter->l[b->level].iter) + if (node_iter != &iter->l[b->c.level].iter) { __bch2_btree_node_iter_fix(iter, b, node_iter, t, - where, clobber_u64s, new_u64s); + where, clobber_u64s, new_u64s); - trans_for_each_iter_with_node(iter->trans, b, linked) + if (bch2_debug_check_iterators) + bch2_btree_node_iter_verify(node_iter, b); + } + + trans_for_each_iter_with_node(iter->trans, b, linked) { __bch2_btree_node_iter_fix(linked, b, - &linked->l[b->level].iter, t, - where, clobber_u64s, new_u64s); + &linked->l[b->c.level].iter, t, + where, clobber_u64s, new_u64s); + bch2_btree_iter_verify_level(linked, b->c.level); + } } static inline struct bkey_s_c __btree_iter_unpack(struct btree_iter *iter, @@ -658,26 +842,45 @@ static inline struct bkey_s_c __btree_iter_unpack(struct btree_iter *iter, ret = bkey_disassemble(l->b, k, u); - if (debug_check_bkeys(iter->trans->c)) + /* + * 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(iter->trans->c, l->b, ret); return ret; } /* peek_all() doesn't skip deleted keys */ -static inline struct bkey_s_c __btree_iter_peek_all(struct btree_iter *iter, - struct btree_iter_level *l, - struct bkey *u) +static inline struct bkey_s_c btree_iter_level_peek_all(struct btree_iter *iter, + struct btree_iter_level *l) { - return __btree_iter_unpack(iter, l, u, + return __btree_iter_unpack(iter, l, &iter->k, bch2_btree_node_iter_peek_all(&l->iter, l->b)); } -static inline struct bkey_s_c __btree_iter_peek(struct btree_iter *iter, - struct btree_iter_level *l) +static inline struct bkey_s_c btree_iter_level_peek(struct btree_iter *iter, + struct btree_iter_level *l) { - return __btree_iter_unpack(iter, l, &iter->k, + struct bkey_s_c k = __btree_iter_unpack(iter, l, &iter->k, bch2_btree_node_iter_peek(&l->iter, l->b)); + + iter->real_pos = k.k ? k.k->p : l->b->key.k.p; + return k; +} + +static inline struct bkey_s_c btree_iter_level_prev(struct btree_iter *iter, + struct btree_iter_level *l) +{ + struct bkey_s_c k = __btree_iter_unpack(iter, l, &iter->k, + bch2_btree_node_iter_prev(&l->iter, l->b)); + + iter->real_pos = k.k ? k.k->p : l->b->data->min_key; + return k; } static inline bool btree_iter_advance_to_pos(struct btree_iter *iter, @@ -688,7 +891,7 @@ static inline bool btree_iter_advance_to_pos(struct btree_iter *iter, int nr_advanced = 0; while ((k = bch2_btree_node_iter_peek_all(&l->iter, l->b)) && - btree_iter_pos_cmp(iter, l->b, k) < 0) { + bkey_iter_pos_cmp(l->b, k, &iter->real_pos) < 0) { if (max_advance > 0 && nr_advanced >= max_advance) return false; @@ -712,7 +915,7 @@ static void btree_iter_verify_new_node(struct btree_iter *iter, struct btree *b) if (!IS_ENABLED(CONFIG_BCACHEFS_DEBUG)) return; - plevel = b->level + 1; + plevel = b->c.level + 1; if (!btree_iter_node(iter, plevel)) return; @@ -726,31 +929,27 @@ static void btree_iter_verify_new_node(struct btree_iter *iter, struct btree *b) if (!k || bkey_deleted(k) || bkey_cmp_left_packed(l->b, k, &b->key.k.p)) { - char buf[100]; + char buf1[100]; + char buf2[100]; + char buf3[100]; + char buf4[100]; struct bkey uk = bkey_unpack_key(b, k); - bch2_bkey_to_text(&PBUF(buf), &uk); - panic("parent iter doesn't point to new node:\n%s\n%llu:%llu\n", - buf, b->key.k.p.inode, b->key.k.p.offset); + bch2_dump_btree_node(iter->trans->c, l->b); + bch2_bpos_to_text(&PBUF(buf1), iter->real_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[iter->btree_id], buf1, + buf2, buf3, buf4); } if (!parent_locked) - btree_node_unlock(iter, b->level + 1); -} - -static inline bool btree_iter_pos_after_node(struct btree_iter *iter, - struct btree *b) -{ - return __btree_iter_pos_cmp(iter, NULL, - bkey_to_packed(&b->key), true) < 0; -} - -static inline bool btree_iter_pos_in_node(struct btree_iter *iter, - struct btree *b) -{ - return iter->btree_id == b->btree_id && - bkey_cmp(iter->pos, b->data->min_key) >= 0 && - !btree_iter_pos_after_node(iter, b); + btree_node_unlock(iter, b->c.level + 1); } static inline void __btree_iter_init(struct btree_iter *iter, @@ -758,12 +957,12 @@ static inline void __btree_iter_init(struct btree_iter *iter, { struct btree_iter_level *l = &iter->l[level]; - bch2_btree_node_iter_init(&l->iter, l->b, &iter->pos); + bch2_btree_node_iter_init(&l->iter, l->b, &iter->real_pos); - if (iter->flags & BTREE_ITER_IS_EXTENTS) - btree_iter_advance_to_pos(iter, l, -1); - - /* Skip to first non whiteout: */ + /* + * Iterators to interior nodes should always be pointed at the first non + * whiteout: + */ if (level) bch2_btree_node_iter_peek(&l->iter, l->b); @@ -773,14 +972,16 @@ static inline void __btree_iter_init(struct btree_iter *iter, static inline void btree_iter_node_set(struct btree_iter *iter, struct btree *b) { + BUG_ON(btree_iter_type(iter) == BTREE_ITER_CACHED); + btree_iter_verify_new_node(iter, b); EBUG_ON(!btree_iter_pos_in_node(iter, b)); - EBUG_ON(b->lock.state.seq & 1); + EBUG_ON(b->c.lock.state.seq & 1); - iter->l[b->level].lock_seq = b->lock.state.seq; - iter->l[b->level].b = b; - __btree_iter_init(iter, b->level); + iter->l[b->c.level].lock_seq = b->c.lock.state.seq; + iter->l[b->c.level].b = b; + __btree_iter_init(iter, b->c.level); } /* @@ -793,35 +994,34 @@ void bch2_btree_iter_node_replace(struct btree_iter *iter, struct btree *b) struct btree_iter *linked; trans_for_each_iter(iter->trans, linked) - if (btree_iter_pos_in_node(linked, b)) { + if (btree_iter_type(linked) != BTREE_ITER_CACHED && + btree_iter_pos_in_node(linked, b)) { /* * bch2_btree_iter_node_drop() has already been called - * the old node we're replacing has already been * unlocked and the pointer invalidated */ - BUG_ON(btree_node_locked(linked, b->level)); + BUG_ON(btree_node_locked(linked, b->c.level)); - t = btree_lock_want(linked, b->level); + t = btree_lock_want(linked, b->c.level); if (t != BTREE_NODE_UNLOCKED) { - six_lock_increment(&b->lock, t); - mark_btree_node_locked(linked, b->level, t); + six_lock_increment(&b->c.lock, t); + mark_btree_node_locked(linked, b->c.level, t); } btree_iter_node_set(linked, b); } - - six_unlock_intent(&b->lock); } void bch2_btree_iter_node_drop(struct btree_iter *iter, struct btree *b) { struct btree_iter *linked; - unsigned level = b->level; + unsigned level = b->c.level; trans_for_each_iter(iter->trans, linked) if (linked->l[level].b == b) { - __btree_node_unlock(linked, level); - linked->l[level].b = BTREE_ITER_NOT_END; + btree_node_unlock(linked, level); + linked->l[level].b = BTREE_ITER_NO_NODE_DROP; } } @@ -834,22 +1034,31 @@ void bch2_btree_iter_reinit_node(struct btree_iter *iter, struct btree *b) struct btree_iter *linked; trans_for_each_iter_with_node(iter->trans, b, linked) - __btree_iter_init(linked, b->level); + __btree_iter_init(linked, b->c.level); +} + +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_iter_lock_root(struct btree_iter *iter, - unsigned depth_want) + unsigned depth_want, + unsigned long trace_ip) { struct bch_fs *c = iter->trans->c; - struct btree *b; + struct btree *b, **rootp = &c->btree_roots[iter->btree_id].b; enum six_lock_type lock_type; unsigned i; EBUG_ON(iter->nodes_locked); while (1) { - b = READ_ONCE(c->btree_roots[iter->btree_id].b); - iter->level = READ_ONCE(b->level); + b = READ_ONCE(*rootp); + iter->level = READ_ONCE(b->c.level); if (unlikely(iter->level < depth_want)) { /* @@ -859,29 +1068,33 @@ static inline int btree_iter_lock_root(struct btree_iter *iter, * that depth */ iter->level = depth_want; - iter->l[iter->level].b = NULL; + for (i = iter->level; i < BTREE_MAX_DEPTH; i++) + iter->l[i].b = NULL; return 1; } lock_type = __btree_lock_want(iter, iter->level); if (unlikely(!btree_node_lock(b, POS_MAX, iter->level, - iter, lock_type, true))) + iter, lock_type, + lock_root_check_fn, rootp, + trace_ip))) return -EINTR; - if (likely(b == c->btree_roots[iter->btree_id].b && - b->level == iter->level && + if (likely(b == READ_ONCE(*rootp) && + b->c.level == iter->level && !race_fault())) { for (i = 0; i < iter->level; i++) - iter->l[i].b = BTREE_ITER_NOT_END; + iter->l[i].b = BTREE_ITER_NO_NODE_LOCK_ROOT; iter->l[iter->level].b = b; + for (i = iter->level + 1; i < BTREE_MAX_DEPTH; i++) + iter->l[i].b = NULL; mark_btree_node_locked(iter, iter->level, lock_type); btree_iter_node_set(iter, b); return 0; - } - six_unlock_type(&b->lock, lock_type); + six_unlock_type(&b->c.lock, lock_type); } } @@ -892,83 +1105,146 @@ static void btree_iter_prefetch(struct btree_iter *iter) struct btree_iter_level *l = &iter->l[iter->level]; struct btree_node_iter node_iter = l->iter; struct bkey_packed *k; - BKEY_PADDED(k) tmp; + struct bkey_buf tmp; unsigned nr = test_bit(BCH_FS_STARTED, &c->flags) ? (iter->level > 1 ? 0 : 2) : (iter->level > 1 ? 1 : 16); bool was_locked = btree_node_locked(iter, iter->level); + bch2_bkey_buf_init(&tmp); + while (nr) { if (!bch2_btree_node_relock(iter, iter->level)) - return; + break; bch2_btree_node_iter_advance(&node_iter, l->b); k = bch2_btree_node_iter_peek(&node_iter, l->b); if (!k) break; - bch2_bkey_unpack(l->b, &tmp.k, k); - bch2_btree_node_prefetch(c, iter, &tmp.k, iter->level - 1); + bch2_bkey_buf_unpack(&tmp, c, l->b, k); + bch2_btree_node_prefetch(c, iter, tmp.k, iter->btree_id, + iter->level - 1); } if (!was_locked) btree_node_unlock(iter, iter->level); + + bch2_bkey_buf_exit(&tmp, c); +} + +static noinline void btree_node_mem_ptr_set(struct btree_iter *iter, + unsigned plevel, struct btree *b) +{ + struct btree_iter_level *l = &iter->l[plevel]; + bool locked = btree_node_locked(iter, plevel); + struct bkey_packed *k; + struct bch_btree_ptr_v2 *bp; + + if (!bch2_btree_node_relock(iter, plevel)) + return; + + k = bch2_btree_node_iter_peek_all(&l->iter, l->b); + BUG_ON(k->type != KEY_TYPE_btree_ptr_v2); + + bp = (void *) bkeyp_val(&l->b->format, k); + bp->mem_ptr = (unsigned long)b; + + if (!locked) + btree_node_unlock(iter, plevel); } -static inline int btree_iter_down(struct btree_iter *iter) +static __always_inline int btree_iter_down(struct btree_iter *iter, + unsigned long trace_ip) { struct bch_fs *c = iter->trans->c; struct btree_iter_level *l = &iter->l[iter->level]; struct btree *b; unsigned level = iter->level - 1; enum six_lock_type lock_type = __btree_lock_want(iter, level); - BKEY_PADDED(k) tmp; + struct bkey_buf tmp; + int ret; - BUG_ON(!btree_node_locked(iter, iter->level)); + EBUG_ON(!btree_node_locked(iter, iter->level)); - bch2_bkey_unpack(l->b, &tmp.k, + bch2_bkey_buf_init(&tmp); + bch2_bkey_buf_unpack(&tmp, c, l->b, bch2_btree_node_iter_peek(&l->iter, l->b)); - b = bch2_btree_node_get(c, iter, &tmp.k, level, lock_type, true); - if (unlikely(IS_ERR(b))) - return PTR_ERR(b); + b = bch2_btree_node_get(c, iter, tmp.k, level, lock_type, trace_ip); + ret = PTR_ERR_OR_ZERO(b); + if (unlikely(ret)) + goto err; mark_btree_node_locked(iter, level, lock_type); btree_iter_node_set(iter, b); + if (tmp.k->k.type == KEY_TYPE_btree_ptr_v2 && + unlikely(b != btree_node_mem_ptr(tmp.k))) + btree_node_mem_ptr_set(iter, level + 1, b); + if (iter->flags & BTREE_ITER_PREFETCH) btree_iter_prefetch(iter); + if (btree_node_read_locked(iter, level + 1)) + btree_node_unlock(iter, level + 1); iter->level = level; - return 0; -} - -static void btree_iter_up(struct btree_iter *iter) -{ - btree_node_unlock(iter, iter->level++); + bch2_btree_iter_verify_locks(iter); +err: + bch2_bkey_buf_exit(&tmp, c); + return ret; } -int __must_check __bch2_btree_iter_traverse(struct btree_iter *); +static int btree_iter_traverse_one(struct btree_iter *, unsigned long); -static int __btree_iter_traverse_all(struct btree_trans *trans, - struct btree_iter *iter, int ret) +static int __btree_iter_traverse_all(struct btree_trans *trans, int ret, + unsigned long trace_ip) { struct bch_fs *c = trans->c; + struct btree_iter *iter; u8 sorted[BTREE_ITER_MAX]; - unsigned i, nr_sorted = 0; + int i, nr_sorted = 0; + bool relock_fail; - trans_for_each_iter(trans, iter) - sorted[nr_sorted++] = iter - trans->iters; + if (trans->in_traverse_all) + return -EINTR; + + trans->in_traverse_all = true; +retry_all: + nr_sorted = 0; + relock_fail = false; + + trans_for_each_iter(trans, iter) { + if (!bch2_btree_iter_relock(iter, _THIS_IP_)) + relock_fail = true; + sorted[nr_sorted++] = iter->idx; + } + + if (!relock_fail) { + trans->in_traverse_all = false; + return 0; + } #define btree_iter_cmp_by_idx(_l, _r) \ - btree_iter_cmp(&trans->iters[_l], &trans->iters[_r]) + btree_iter_lock_cmp(&trans->iters[_l], &trans->iters[_r]) bubble_sort(sorted, nr_sorted, btree_iter_cmp_by_idx); #undef btree_iter_cmp_by_idx -retry_all: - bch2_btree_trans_unlock(trans); + for (i = nr_sorted - 2; i >= 0; --i) { + struct btree_iter *iter1 = trans->iters + sorted[i]; + struct btree_iter *iter2 = trans->iters + sorted[i + 1]; + + if (iter1->btree_id == iter2->btree_id && + iter1->locks_want < iter2->locks_want) + __bch2_btree_iter_upgrade(iter1, iter2->locks_want); + else if (!iter1->locks_want && iter2->locks_want) + __bch2_btree_iter_upgrade(iter1, 1); + } + + bch2_trans_unlock(trans); + cond_resched(); if (unlikely(ret == -ENOMEM)) { struct closure cl; @@ -983,8 +1259,6 @@ retry_all: if (unlikely(ret == -EIO)) { trans->error = true; - iter->flags |= BTREE_ITER_ERROR; - iter->l[iter->level].b = BTREE_ITER_NOT_END; goto out; } @@ -992,39 +1266,65 @@ retry_all: /* Now, redo traversals in correct order: */ for (i = 0; i < nr_sorted; i++) { - iter = &trans->iters[sorted[i]]; + unsigned idx = sorted[i]; - do { - ret = __bch2_btree_iter_traverse(iter); - } while (ret == -EINTR); + /* + * sucessfully traversing one iterator can cause another to be + * unlinked, in btree_key_cache_fill() + */ + if (!(trans->iters_linked & (1ULL << idx))) + continue; + ret = btree_iter_traverse_one(&trans->iters[idx], _THIS_IP_); if (ret) goto retry_all; } - ret = btree_trans_has_multiple_iters(trans) ? -EINTR : 0; + if (hweight64(trans->iters_live) > 1) + ret = -EINTR; + else + trans_for_each_iter(trans, iter) + if (iter->flags & BTREE_ITER_KEEP_UNTIL_COMMIT) { + ret = -EINTR; + break; + } out: bch2_btree_cache_cannibalize_unlock(c); + + trans->in_traverse_all = false; + + trace_trans_traverse_all(trans->ip, trace_ip); return ret; } int bch2_btree_iter_traverse_all(struct btree_trans *trans) { - return __btree_iter_traverse_all(trans, NULL, 0); + return __btree_iter_traverse_all(trans, 0, _RET_IP_); } -static unsigned btree_iter_up_until_locked(struct btree_iter *iter, - bool check_pos) +static inline bool btree_iter_good_node(struct btree_iter *iter, + unsigned l, int check_pos) +{ + if (!is_btree_node(iter, l) || + !bch2_btree_node_relock(iter, l)) + return false; + + if (check_pos < 0 && btree_iter_pos_before_node(iter, iter->l[l].b)) + return false; + if (check_pos > 0 && btree_iter_pos_after_node(iter, iter->l[l].b)) + return false; + return true; +} + +static inline unsigned btree_iter_up_until_good_node(struct btree_iter *iter, + int check_pos) { unsigned l = iter->level; while (btree_iter_node(iter, l) && - !(is_btree_node(iter, l) && - bch2_btree_node_relock(iter, l) && - (!check_pos || - btree_iter_pos_in_node(iter, iter->l[l].b)))) { + !btree_iter_good_node(iter, l, check_pos)) { btree_node_unlock(iter, l); - iter->l[l].b = BTREE_ITER_NOT_END; + iter->l[l].b = BTREE_ITER_NO_NODE_UP; l++; } @@ -1038,32 +1338,34 @@ static unsigned btree_iter_up_until_locked(struct btree_iter *iter, * Returns 0 on success, -EIO on error (error reading in a btree node). * * On error, caller (peek_node()/peek_key()) must return NULL; the error is - * stashed in the iterator and returned from bch2_btree_iter_unlock(). + * stashed in the iterator and returned from bch2_trans_exit(). */ -int __must_check __bch2_btree_iter_traverse(struct btree_iter *iter) +static int btree_iter_traverse_one(struct btree_iter *iter, + unsigned long trace_ip) { unsigned depth_want = iter->level; - - if (unlikely(iter->level >= BTREE_MAX_DEPTH)) - return 0; - - if (bch2_btree_iter_relock(iter)) - return 0; + int ret = 0; /* - * XXX: correctly using BTREE_ITER_UPTODATE should make using check_pos - * here unnecessary + * if we need interior nodes locked, call btree_iter_relock() to make + * sure we walk back up enough that we lock them: */ - iter->level = btree_iter_up_until_locked(iter, true); + if (iter->uptodate == BTREE_ITER_NEED_RELOCK || + iter->locks_want > 1) + bch2_btree_iter_relock(iter, _THIS_IP_); - /* - * If we've got a btree node locked (i.e. we aren't about to relock the - * root) - advance its node iterator if necessary: - * - * XXX correctly using BTREE_ITER_UPTODATE should make this unnecessary - */ - if (btree_iter_node(iter, iter->level)) - btree_iter_advance_to_pos(iter, &iter->l[iter->level], -1); + if (btree_iter_type(iter) == BTREE_ITER_CACHED) { + ret = bch2_btree_iter_traverse_cached(iter); + goto out; + } + + if (iter->uptodate < BTREE_ITER_NEED_RELOCK) + goto out; + + if (unlikely(iter->level >= BTREE_MAX_DEPTH)) + goto out; + + iter->level = btree_iter_up_until_good_node(iter, 0); /* * Note: iter->nodes[iter->level] may be temporarily NULL here - that @@ -1072,48 +1374,83 @@ int __must_check __bch2_btree_iter_traverse(struct btree_iter *iter) * btree_iter_lock_root() comes next and that it can't fail */ while (iter->level > depth_want) { - int ret = btree_iter_node(iter, iter->level) - ? btree_iter_down(iter) - : btree_iter_lock_root(iter, depth_want); + ret = btree_iter_node(iter, iter->level) + ? btree_iter_down(iter, trace_ip) + : btree_iter_lock_root(iter, depth_want, trace_ip); if (unlikely(ret)) { - if (ret == 1) - return 0; + if (ret == 1) { + /* + * Got to the end of the btree (in + * BTREE_ITER_NODES mode) + */ + ret = 0; + goto out; + } iter->level = depth_want; - iter->l[iter->level].b = BTREE_ITER_NOT_END; - return ret; + + if (ret == -EIO) { + iter->flags |= BTREE_ITER_ERROR; + iter->l[iter->level].b = + BTREE_ITER_NO_NODE_ERROR; + } else { + iter->l[iter->level].b = + BTREE_ITER_NO_NODE_DOWN; + } + goto out; } } iter->uptodate = BTREE_ITER_NEED_PEEK; - - bch2_btree_trans_verify_locks(iter->trans); - __bch2_btree_iter_verify(iter, iter->l[iter->level].b); - return 0; +out: + trace_iter_traverse(iter->trans->ip, trace_ip, + iter->btree_id, &iter->real_pos, ret); + bch2_btree_iter_verify(iter); + return ret; } -int __must_check bch2_btree_iter_traverse(struct btree_iter *iter) +static int __must_check __bch2_btree_iter_traverse(struct btree_iter *iter) { + struct btree_trans *trans = iter->trans; int ret; - ret = __bch2_btree_iter_traverse(iter); + ret = bch2_trans_cond_resched(trans) ?: + btree_iter_traverse_one(iter, _RET_IP_); if (unlikely(ret)) - ret = __btree_iter_traverse_all(iter->trans, iter, ret); - - BUG_ON(ret == -EINTR && !btree_trans_has_multiple_iters(iter->trans)); + ret = __btree_iter_traverse_all(trans, ret, _RET_IP_); return ret; } -static inline void bch2_btree_iter_checks(struct btree_iter *iter, - enum btree_iter_type type) +/* + * Note: + * bch2_btree_iter_traverse() is for external users, btree_iter_traverse() is + * for internal btree iterator users + * + * bch2_btree_iter_traverse sets iter->real_pos to iter->pos, + * btree_iter_traverse() does not: + */ +static inline int __must_check +btree_iter_traverse(struct btree_iter *iter) { - EBUG_ON(iter->btree_id >= BTREE_ID_NR); - EBUG_ON(!!(iter->flags & BTREE_ITER_IS_EXTENTS) != - (iter->btree_id == BTREE_ID_EXTENTS && - type != BTREE_ITER_NODES)); + return iter->uptodate >= BTREE_ITER_NEED_RELOCK + ? __bch2_btree_iter_traverse(iter) + : 0; +} - bch2_btree_trans_verify_locks(iter->trans); +int __must_check +bch2_btree_iter_traverse(struct btree_iter *iter) +{ + int ret; + + btree_iter_set_search_pos(iter, btree_iter_search_key(iter)); + + ret = btree_iter_traverse(iter); + if (ret) + return ret; + + iter->should_be_locked = true; + return 0; } /* Iterate across nodes (leaf and interior nodes) */ @@ -1123,12 +1460,10 @@ struct btree *bch2_btree_iter_peek_node(struct btree_iter *iter) struct btree *b; int ret; - bch2_btree_iter_checks(iter, BTREE_ITER_NODES); + EBUG_ON(btree_iter_type(iter) != BTREE_ITER_NODES); + bch2_btree_iter_verify(iter); - if (iter->uptodate == BTREE_ITER_UPTODATE) - return iter->l[iter->level].b; - - ret = bch2_btree_iter_traverse(iter); + ret = btree_iter_traverse(iter); if (ret) return NULL; @@ -1136,20 +1471,23 @@ struct btree *bch2_btree_iter_peek_node(struct btree_iter *iter) if (!b) return NULL; - BUG_ON(bkey_cmp(b->key.k.p, iter->pos) < 0); + BUG_ON(bpos_cmp(b->key.k.p, iter->pos) < 0); + + iter->pos = iter->real_pos = b->key.k.p; - iter->pos = b->key.k.p; - iter->uptodate = BTREE_ITER_UPTODATE; + bch2_btree_iter_verify(iter); + iter->should_be_locked = true; return b; } -struct btree *bch2_btree_iter_next_node(struct btree_iter *iter, unsigned depth) +struct btree *bch2_btree_iter_next_node(struct btree_iter *iter) { struct btree *b; int ret; - bch2_btree_iter_checks(iter, BTREE_ITER_NODES); + EBUG_ON(btree_iter_type(iter) != BTREE_ITER_NODES); + bch2_btree_iter_verify(iter); /* already got to end? */ if (!btree_iter_node(iter, iter->level)) @@ -1157,12 +1495,12 @@ struct btree *bch2_btree_iter_next_node(struct btree_iter *iter, unsigned depth) bch2_trans_cond_resched(iter->trans); - btree_iter_up(iter); + btree_node_unlock(iter, iter->level); + iter->l[iter->level].b = BTREE_ITER_NO_NODE_UP; + iter->level++; - if (!bch2_btree_node_relock(iter, iter->level)) - btree_iter_set_dirty(iter, BTREE_ITER_NEED_RELOCK); - - ret = bch2_btree_iter_traverse(iter); + btree_iter_set_dirty(iter, BTREE_ITER_NEED_TRAVERSE); + ret = btree_iter_traverse(iter); if (ret) return NULL; @@ -1171,75 +1509,59 @@ struct btree *bch2_btree_iter_next_node(struct btree_iter *iter, unsigned depth) if (!b) return NULL; - if (bkey_cmp(iter->pos, b->key.k.p) < 0) { + if (bpos_cmp(iter->pos, b->key.k.p) < 0) { /* * Haven't gotten to the end of the parent node: go back down to * the next child node */ + btree_iter_set_search_pos(iter, bpos_successor(iter->pos)); - /* - * We don't really want to be unlocking here except we can't - * directly tell btree_iter_traverse() "traverse to this level" - * except by setting iter->level, so we have to unlock so we - * don't screw up our lock invariants: - */ - if (btree_node_read_locked(iter, iter->level)) - btree_node_unlock(iter, iter->level); - - /* ick: */ - iter->pos = iter->btree_id == BTREE_ID_INODES - ? btree_type_successor(iter->btree_id, iter->pos) - : bkey_successor(iter->pos); - iter->level = depth; + /* Unlock to avoid screwing up our lock invariants: */ + btree_node_unlock(iter, iter->level); + iter->level = iter->min_depth; btree_iter_set_dirty(iter, BTREE_ITER_NEED_TRAVERSE); - ret = bch2_btree_iter_traverse(iter); + bch2_btree_iter_verify(iter); + + ret = btree_iter_traverse(iter); if (ret) return NULL; b = iter->l[iter->level].b; } - iter->pos = b->key.k.p; - iter->uptodate = BTREE_ITER_UPTODATE; + iter->pos = iter->real_pos = b->key.k.p; + + bch2_btree_iter_verify(iter); + iter->should_be_locked = true; return b; } /* Iterate across keys (in leaf nodes only) */ -void bch2_btree_iter_set_pos_same_leaf(struct btree_iter *iter, struct bpos new_pos) +static void btree_iter_set_search_pos(struct btree_iter *iter, struct bpos new_pos) { - struct btree_iter_level *l = &iter->l[0]; - - EBUG_ON(iter->level != 0); - EBUG_ON(bkey_cmp(new_pos, iter->pos) < 0); - EBUG_ON(!btree_node_locked(iter, 0)); - EBUG_ON(bkey_cmp(new_pos, l->b->key.k.p) > 0); + struct bpos old_pos = iter->real_pos; + int cmp = bpos_cmp(new_pos, iter->real_pos); + unsigned l = iter->level; - iter->pos = new_pos; - btree_iter_set_dirty(iter, BTREE_ITER_NEED_PEEK); + if (!cmp) + goto out; - btree_iter_advance_to_pos(iter, l, -1); + iter->real_pos = new_pos; + iter->should_be_locked = false; - if (bch2_btree_node_iter_end(&l->iter) && - btree_iter_pos_after_node(iter, l->b)) + if (unlikely(btree_iter_type(iter) == BTREE_ITER_CACHED)) { + btree_node_unlock(iter, 0); + iter->l[0].b = BTREE_ITER_NO_NODE_CACHED; btree_iter_set_dirty(iter, BTREE_ITER_NEED_TRAVERSE); -} - -void bch2_btree_iter_set_pos(struct btree_iter *iter, struct bpos new_pos) -{ - int cmp = bkey_cmp(new_pos, iter->pos); - unsigned level; - - if (!cmp) return; + } - iter->pos = new_pos; - - level = btree_iter_up_until_locked(iter, true); + l = btree_iter_up_until_good_node(iter, cmp); - if (btree_iter_node(iter, level)) { + if (btree_iter_node(iter, l)) { /* * 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 @@ -1247,546 +1569,632 @@ void bch2_btree_iter_set_pos(struct btree_iter *iter, struct bpos new_pos) * is expensive). */ if (cmp < 0 || - !btree_iter_advance_to_pos(iter, &iter->l[level], 8)) - __btree_iter_init(iter, level); + !btree_iter_advance_to_pos(iter, &iter->l[l], 8)) + __btree_iter_init(iter, l); /* Don't leave it locked if we're not supposed to: */ - if (btree_lock_want(iter, level) == BTREE_NODE_UNLOCKED) - btree_node_unlock(iter, level); + if (btree_lock_want(iter, l) == BTREE_NODE_UNLOCKED) + btree_node_unlock(iter, l); } - - if (level != iter->level) +out: + if (l != iter->level) btree_iter_set_dirty(iter, BTREE_ITER_NEED_TRAVERSE); else btree_iter_set_dirty(iter, BTREE_ITER_NEED_PEEK); + + bch2_btree_iter_verify(iter); +#ifdef CONFIG_BCACHEFS_DEBUG + trace_iter_set_search_pos(iter->trans->ip, _RET_IP_, + iter->btree_id, + &old_pos, &new_pos, l); +#endif } -static inline struct bkey_s_c btree_iter_peek_uptodate(struct btree_iter *iter) +inline bool bch2_btree_iter_advance(struct btree_iter *iter) { - struct btree_iter_level *l = &iter->l[0]; - struct bkey_s_c ret = { .k = &iter->k }; - - if (!bkey_deleted(&iter->k)) { - EBUG_ON(bch2_btree_node_iter_end(&l->iter)); - ret.v = bkeyp_val(&l->b->format, - __bch2_btree_node_iter_peek_all(&l->iter, l->b)); - } + struct bpos pos = iter->k.p; + bool ret = bpos_cmp(pos, POS_MAX) != 0; - if (debug_check_bkeys(iter->trans->c) && - !bkey_deleted(ret.k)) - bch2_bkey_debugcheck(iter->trans->c, l->b, ret); + if (ret && !(iter->flags & BTREE_ITER_IS_EXTENTS)) + pos = bkey_successor(iter, pos); + bch2_btree_iter_set_pos(iter, pos); return ret; } -struct bkey_s_c bch2_btree_iter_peek(struct btree_iter *iter) +inline bool bch2_btree_iter_rewind(struct btree_iter *iter) { - struct btree_iter_level *l = &iter->l[0]; - struct bkey_s_c k; - int ret; + struct bpos pos = bkey_start_pos(&iter->k); + bool ret = bpos_cmp(pos, POS_MIN) != 0; - bch2_btree_iter_checks(iter, BTREE_ITER_KEYS); - - if (iter->uptodate == BTREE_ITER_UPTODATE) - return btree_iter_peek_uptodate(iter); - - while (1) { - ret = bch2_btree_iter_traverse(iter); - if (unlikely(ret)) - return bkey_s_c_err(ret); - - k = __btree_iter_peek(iter, l); - if (likely(k.k)) - break; - - /* got to the end of the leaf, iterator needs to be traversed: */ - iter->pos = l->b->key.k.p; - iter->uptodate = BTREE_ITER_NEED_TRAVERSE; - - if (!bkey_cmp(iter->pos, POS_MAX)) - return bkey_s_c_null; + if (ret && !(iter->flags & BTREE_ITER_IS_EXTENTS)) + pos = bkey_predecessor(iter, pos); + bch2_btree_iter_set_pos(iter, pos); + return ret; +} - iter->pos = btree_type_successor(iter->btree_id, iter->pos); - } +static inline bool btree_iter_set_pos_to_next_leaf(struct btree_iter *iter) +{ + struct bpos next_pos = iter->l[0].b->key.k.p; + bool ret = bpos_cmp(next_pos, POS_MAX) != 0; /* - * iter->pos should always be equal to the key we just - * returned - except extents can straddle iter->pos: + * Typically, we don't want to modify iter->pos here, since that + * indicates where we searched from - unless we got to the end of the + * btree, in that case we want iter->pos to reflect that: */ - if (!(iter->flags & BTREE_ITER_IS_EXTENTS) || - bkey_cmp(bkey_start_pos(k.k), iter->pos) > 0) - iter->pos = bkey_start_pos(k.k); + if (ret) + btree_iter_set_search_pos(iter, bpos_successor(next_pos)); + else + bch2_btree_iter_set_pos(iter, POS_MAX); - iter->uptodate = BTREE_ITER_UPTODATE; - return k; + return ret; } -static noinline -struct bkey_s_c bch2_btree_iter_peek_next_leaf(struct btree_iter *iter) +static inline bool btree_iter_set_pos_to_prev_leaf(struct btree_iter *iter) { - struct btree_iter_level *l = &iter->l[0]; - - iter->pos = l->b->key.k.p; - iter->uptodate = BTREE_ITER_NEED_TRAVERSE; - - if (!bkey_cmp(iter->pos, POS_MAX)) - return bkey_s_c_null; + struct bpos next_pos = iter->l[0].b->data->min_key; + bool ret = bpos_cmp(next_pos, POS_MIN) != 0; - iter->pos = btree_type_successor(iter->btree_id, iter->pos); + if (ret) + btree_iter_set_search_pos(iter, bpos_predecessor(next_pos)); + else + bch2_btree_iter_set_pos(iter, POS_MIN); - return bch2_btree_iter_peek(iter); + return ret; } -struct bkey_s_c bch2_btree_iter_next(struct btree_iter *iter) +static inline struct bkey_i *btree_trans_peek_updates(struct btree_iter *iter, + struct bpos pos) { - struct btree_iter_level *l = &iter->l[0]; - struct bkey_packed *p; - struct bkey_s_c k; + struct btree_insert_entry *i; - bch2_btree_iter_checks(iter, BTREE_ITER_KEYS); - - if (unlikely(iter->uptodate != BTREE_ITER_UPTODATE)) { - k = bch2_btree_iter_peek(iter); - if (IS_ERR_OR_NULL(k.k)) - return k; - } - - do { - bch2_btree_node_iter_advance(&l->iter, l->b); - p = bch2_btree_node_iter_peek_all(&l->iter, l->b); - if (unlikely(!p)) - return bch2_btree_iter_peek_next_leaf(iter); - } while (bkey_whiteout(p)); + if (!(iter->flags & BTREE_ITER_WITH_UPDATES)) + return NULL; - k = __btree_iter_unpack(iter, l, &iter->k, p); + trans_for_each_update(iter->trans, i) + if ((cmp_int(iter->btree_id, i->iter->btree_id) ?: + bkey_cmp(pos, i->k->k.p)) <= 0) { + if (iter->btree_id == i->iter->btree_id) + return i->k; + break; + } - EBUG_ON(bkey_cmp(bkey_start_pos(k.k), iter->pos) < 0); - iter->pos = bkey_start_pos(k.k); - return k; + return NULL; } -struct bkey_s_c bch2_btree_iter_prev(struct btree_iter *iter) +/** + * bch2_btree_iter_peek: returns first key greater than or equal to iterator's + * current position + */ +struct bkey_s_c bch2_btree_iter_peek(struct btree_iter *iter) { - struct btree_iter_level *l = &iter->l[0]; - struct bkey_packed *p; + struct bpos search_key = btree_iter_search_key(iter); + struct bkey_i *next_update; struct bkey_s_c k; int ret; - bch2_btree_iter_checks(iter, BTREE_ITER_KEYS); - - if (unlikely(iter->uptodate != BTREE_ITER_UPTODATE)) { - k = bch2_btree_iter_peek(iter); - if (IS_ERR(k.k)) - return k; - } + EBUG_ON(btree_iter_type(iter) != BTREE_ITER_KEYS); + bch2_btree_iter_verify(iter); + bch2_btree_iter_verify_entry_exit(iter); +start: + next_update = btree_trans_peek_updates(iter, search_key); + btree_iter_set_search_pos(iter, search_key); while (1) { - p = bch2_btree_node_iter_prev(&l->iter, l->b); - if (likely(p)) - break; + ret = btree_iter_traverse(iter); + if (unlikely(ret)) + return bkey_s_c_err(ret); - iter->pos = l->b->data->min_key; - if (!bkey_cmp(iter->pos, POS_MIN)) - return bkey_s_c_null; + k = btree_iter_level_peek(iter, &iter->l[0]); - bch2_btree_iter_set_pos(iter, - btree_type_predecessor(iter->btree_id, iter->pos)); + if (next_update && + bpos_cmp(next_update->k.p, iter->real_pos) <= 0) { + iter->k = next_update->k; + k = bkey_i_to_s_c(next_update); + } - ret = bch2_btree_iter_traverse(iter); - if (unlikely(ret)) - return bkey_s_c_err(ret); + if (likely(k.k)) { + if (bkey_deleted(k.k)) { + search_key = bkey_successor(iter, k.k->p); + goto start; + } - p = bch2_btree_node_iter_peek(&l->iter, l->b); - if (p) break; - } + } - k = __btree_iter_unpack(iter, l, &iter->k, p); + if (!btree_iter_set_pos_to_next_leaf(iter)) + return bkey_s_c_null; + } - EBUG_ON(bkey_cmp(bkey_start_pos(k.k), iter->pos) > 0); + /* + * 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 = bkey_start_pos(k.k); - iter->uptodate = BTREE_ITER_UPTODATE; + bch2_btree_iter_verify_entry_exit(iter); + bch2_btree_iter_verify(iter); + iter->should_be_locked = true; return k; } -static inline struct bkey_s_c -__bch2_btree_iter_peek_slot_extents(struct btree_iter *iter) +/** + * bch2_btree_iter_next: returns first key greater than iterator's current + * position + */ +struct bkey_s_c bch2_btree_iter_next(struct btree_iter *iter) +{ + if (!bch2_btree_iter_advance(iter)) + return bkey_s_c_null; + + return bch2_btree_iter_peek(iter); +} + +/** + * bch2_btree_iter_peek_prev: returns first key less than or equal to + * iterator's current position + */ +struct bkey_s_c bch2_btree_iter_peek_prev(struct btree_iter *iter) { struct btree_iter_level *l = &iter->l[0]; - struct btree_node_iter node_iter; struct bkey_s_c k; - struct bkey n; int ret; -recheck: - while ((k = __btree_iter_peek_all(iter, l, &iter->k)).k && - bkey_deleted(k.k) && - bkey_cmp(bkey_start_pos(k.k), iter->pos) == 0) - bch2_btree_node_iter_advance(&l->iter, l->b); + EBUG_ON(btree_iter_type(iter) != BTREE_ITER_KEYS); + EBUG_ON(iter->flags & BTREE_ITER_WITH_UPDATES); + bch2_btree_iter_verify(iter); + bch2_btree_iter_verify_entry_exit(iter); - /* - * iterator is now at the correct position for inserting at iter->pos, - * but we need to keep iterating until we find the first non whiteout so - * we know how big a hole we have, if any: - */ + btree_iter_set_search_pos(iter, iter->pos); + + while (1) { + ret = btree_iter_traverse(iter); + if (unlikely(ret)) { + k = bkey_s_c_err(ret); + goto no_key; + } + + k = btree_iter_level_peek(iter, l); + if (!k.k || + ((iter->flags & BTREE_ITER_IS_EXTENTS) + ? bkey_cmp(bkey_start_pos(k.k), iter->pos) >= 0 + : bkey_cmp(bkey_start_pos(k.k), iter->pos) > 0)) + k = btree_iter_level_prev(iter, l); + + if (likely(k.k)) + break; + + if (!btree_iter_set_pos_to_prev_leaf(iter)) { + k = bkey_s_c_null; + goto no_key; + } + } - node_iter = l->iter; - if (k.k && bkey_whiteout(k.k)) - k = __btree_iter_unpack(iter, l, &iter->k, - bch2_btree_node_iter_peek(&node_iter, l->b)); + EBUG_ON(bkey_cmp(bkey_start_pos(k.k), iter->pos) > 0); + /* Extents can straddle iter->pos: */ + if (bkey_cmp(k.k->p, iter->pos) < 0) + iter->pos = k.k->p; +out: + bch2_btree_iter_verify_entry_exit(iter); + bch2_btree_iter_verify(iter); + iter->should_be_locked = true; + return k; +no_key: /* - * If we got to the end of the node, check if we need to traverse to the - * next node: + * btree_iter_level_peek() may have set iter->k to a key we didn't want, and + * then we errored going to the previous leaf - make sure it's + * consistent with iter->pos: */ - if (unlikely(!k.k && btree_iter_pos_after_node(iter, l->b))) { - btree_iter_set_dirty(iter, BTREE_ITER_NEED_TRAVERSE); - ret = bch2_btree_iter_traverse(iter); - if (unlikely(ret)) - return bkey_s_c_err(ret); + bkey_init(&iter->k); + iter->k.p = iter->pos; + goto out; +} - goto recheck; - } +/** + * bch2_btree_iter_prev: returns first key less than iterator's current + * position + */ +struct bkey_s_c bch2_btree_iter_prev(struct btree_iter *iter) +{ + if (!bch2_btree_iter_rewind(iter)) + return bkey_s_c_null; - if (k.k && - !bkey_whiteout(k.k) && - bkey_cmp(bkey_start_pos(k.k), iter->pos) <= 0) { - /* - * if we skipped forward to find the first non whiteout and - * there _wasn't_ actually a hole, we want the iterator to be - * pointed at the key we found: - */ - l->iter = node_iter; + return bch2_btree_iter_peek_prev(iter); +} - EBUG_ON(bkey_cmp(k.k->p, iter->pos) < 0); - EBUG_ON(bkey_deleted(k.k)); - iter->uptodate = BTREE_ITER_UPTODATE; - return k; - } +struct bkey_s_c bch2_btree_iter_peek_slot(struct btree_iter *iter) +{ + struct bpos search_key; + struct bkey_s_c k; + int ret; - /* hole */ + EBUG_ON(btree_iter_type(iter) != BTREE_ITER_KEYS && + btree_iter_type(iter) != BTREE_ITER_CACHED); + bch2_btree_iter_verify(iter); + bch2_btree_iter_verify_entry_exit(iter); - /* holes can't span inode numbers: */ - if (iter->pos.offset == KEY_OFFSET_MAX) { + /* extents can't span inode numbers: */ + if ((iter->flags & BTREE_ITER_IS_EXTENTS) && + unlikely(iter->pos.offset == KEY_OFFSET_MAX)) { if (iter->pos.inode == KEY_INODE_MAX) return bkey_s_c_null; - iter->pos = bkey_successor(iter->pos); - goto recheck; + bch2_btree_iter_set_pos(iter, bpos_nosnap_successor(iter->pos)); } - if (!k.k) - k.k = &l->b->key.k; + search_key = btree_iter_search_key(iter); + btree_iter_set_search_pos(iter, search_key); - bkey_init(&n); - n.p = iter->pos; - bch2_key_resize(&n, - min_t(u64, KEY_SIZE_MAX, - (k.k->p.inode == n.p.inode - ? bkey_start_offset(k.k) - : KEY_OFFSET_MAX) - - n.p.offset)); + ret = btree_iter_traverse(iter); + if (unlikely(ret)) + return bkey_s_c_err(ret); - EBUG_ON(!n.size); + if (btree_iter_type(iter) == BTREE_ITER_CACHED || + !(iter->flags & BTREE_ITER_IS_EXTENTS)) { + struct bkey_i *next_update; + struct bkey_cached *ck; - iter->k = n; - iter->uptodate = BTREE_ITER_UPTODATE; - return (struct bkey_s_c) { &iter->k, NULL }; -} + switch (btree_iter_type(iter)) { + case BTREE_ITER_KEYS: + k = btree_iter_level_peek_all(iter, &iter->l[0]); + EBUG_ON(k.k && bkey_deleted(k.k) && bkey_cmp(k.k->p, iter->pos) == 0); + break; + case BTREE_ITER_CACHED: + ck = (void *) iter->l[0].b; + EBUG_ON(iter->btree_id != ck->key.btree_id || + bkey_cmp(iter->pos, ck->key.pos)); + BUG_ON(!ck->valid); -static inline struct bkey_s_c -__bch2_btree_iter_peek_slot(struct btree_iter *iter) -{ - struct btree_iter_level *l = &iter->l[0]; - struct bkey_s_c k; - int ret; + k = bkey_i_to_s_c(ck->k); + break; + case BTREE_ITER_NODES: + BUG(); + } - if (iter->flags & BTREE_ITER_IS_EXTENTS) - return __bch2_btree_iter_peek_slot_extents(iter); + next_update = btree_trans_peek_updates(iter, search_key); + if (next_update && + (!k.k || bpos_cmp(next_update->k.p, k.k->p) <= 0)) { + iter->k = next_update->k; + k = bkey_i_to_s_c(next_update); + } + } else { + if ((iter->flags & BTREE_ITER_INTENT)) { + struct btree_iter *child = + btree_iter_child_alloc(iter, _THIS_IP_); -recheck: - while ((k = __btree_iter_peek_all(iter, l, &iter->k)).k && - bkey_deleted(k.k) && - bkey_cmp(k.k->p, iter->pos) == 0) - bch2_btree_node_iter_advance(&l->iter, l->b); + btree_iter_copy(child, iter); + k = bch2_btree_iter_peek(child); - /* - * If we got to the end of the node, check if we need to traverse to the - * next node: - */ - if (unlikely(!k.k && btree_iter_pos_after_node(iter, l->b))) { - btree_iter_set_dirty(iter, BTREE_ITER_NEED_TRAVERSE); - ret = bch2_btree_iter_traverse(iter); - if (unlikely(ret)) - return bkey_s_c_err(ret); + if (k.k && !bkey_err(k)) + iter->k = child->k; + } else { + struct bpos pos = iter->pos; + + k = bch2_btree_iter_peek(iter); + iter->pos = pos; + } - goto recheck; + if (unlikely(bkey_err(k))) + return k; } - if (k.k && - !bkey_deleted(k.k) && - !bkey_cmp(iter->pos, k.k->p)) { - iter->uptodate = BTREE_ITER_UPTODATE; - return k; + if (!(iter->flags & BTREE_ITER_IS_EXTENTS)) { + if (!k.k || + ((iter->flags & BTREE_ITER_ALL_SNAPSHOTS) + ? bpos_cmp(iter->pos, k.k->p) + : bkey_cmp(iter->pos, k.k->p))) { + bkey_init(&iter->k); + iter->k.p = iter->pos; + k = (struct bkey_s_c) { &iter->k, NULL }; + } } else { - /* hole */ - bkey_init(&iter->k); - iter->k.p = iter->pos; - - iter->uptodate = BTREE_ITER_UPTODATE; - return (struct bkey_s_c) { &iter->k, NULL }; + struct bpos next = k.k ? bkey_start_pos(k.k) : POS_MAX; + + if (bkey_cmp(iter->pos, next) < 0) { + 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)); + + k = (struct bkey_s_c) { &iter->k, NULL }; + EBUG_ON(!k.k->size); + } } -} - -struct bkey_s_c bch2_btree_iter_peek_slot(struct btree_iter *iter) -{ - int ret; - - bch2_btree_iter_checks(iter, BTREE_ITER_SLOTS); - if (iter->uptodate == BTREE_ITER_UPTODATE) - return btree_iter_peek_uptodate(iter); + bch2_btree_iter_verify_entry_exit(iter); + bch2_btree_iter_verify(iter); + iter->should_be_locked = true; - ret = bch2_btree_iter_traverse(iter); - if (unlikely(ret)) - return bkey_s_c_err(ret); - - return __bch2_btree_iter_peek_slot(iter); + return k; } struct bkey_s_c bch2_btree_iter_next_slot(struct btree_iter *iter) { - bch2_btree_iter_checks(iter, BTREE_ITER_SLOTS); - - iter->pos = btree_type_successor(iter->btree_id, iter->k.p); - - if (unlikely(iter->uptodate != BTREE_ITER_UPTODATE)) { - /* - * XXX: when we just need to relock we should be able to avoid - * calling traverse, but we need to kill BTREE_ITER_NEED_PEEK - * for that to work - */ - btree_iter_set_dirty(iter, BTREE_ITER_NEED_TRAVERSE); - - return bch2_btree_iter_peek_slot(iter); - } + if (!bch2_btree_iter_advance(iter)) + return bkey_s_c_null; - if (!bkey_deleted(&iter->k)) - bch2_btree_node_iter_advance(&iter->l[0].iter, iter->l[0].b); + return bch2_btree_iter_peek_slot(iter); +} - btree_iter_set_dirty(iter, BTREE_ITER_NEED_PEEK); +struct bkey_s_c bch2_btree_iter_prev_slot(struct btree_iter *iter) +{ + if (!bch2_btree_iter_rewind(iter)) + return bkey_s_c_null; - return __bch2_btree_iter_peek_slot(iter); + return bch2_btree_iter_peek_slot(iter); } static inline void bch2_btree_iter_init(struct btree_trans *trans, - struct btree_iter *iter, enum btree_id btree_id, - struct bpos pos, unsigned flags) + struct btree_iter *iter, enum btree_id btree_id) { struct bch_fs *c = trans->c; unsigned i; - if (btree_id == BTREE_ID_EXTENTS && - !(flags & BTREE_ITER_NODES)) - flags |= BTREE_ITER_IS_EXTENTS; - iter->trans = trans; - iter->pos = pos; - bkey_init(&iter->k); - iter->k.p = pos; - iter->flags = flags; iter->uptodate = BTREE_ITER_NEED_TRAVERSE; iter->btree_id = btree_id; + iter->real_pos = POS_MIN; iter->level = 0; - iter->locks_want = flags & BTREE_ITER_INTENT ? 1 : 0; + iter->min_depth = 0; + iter->locks_want = 0; iter->nodes_locked = 0; iter->nodes_intent_locked = 0; for (i = 0; i < ARRAY_SIZE(iter->l); i++) - iter->l[i].b = NULL; - iter->l[iter->level].b = BTREE_ITER_NOT_END; + iter->l[i].b = BTREE_ITER_NO_NODE_INIT; prefetch(c->btree_roots[btree_id].b); } /* new transactional stuff: */ -int bch2_trans_iter_put(struct btree_trans *trans, - struct btree_iter *iter) +static void btree_iter_child_free(struct btree_iter *iter) { - int ret = btree_iter_err(iter); + struct btree_iter *child = btree_iter_child(iter); - trans->iters_live &= ~(1ULL << iter->idx); - return ret; + if (child) { + bch2_trans_iter_free(iter->trans, child); + iter->child_idx = U8_MAX; + } +} + +static struct btree_iter *btree_iter_child_alloc(struct btree_iter *iter, + unsigned long ip) +{ + struct btree_trans *trans = iter->trans; + struct btree_iter *child = btree_iter_child(iter); + + if (!child) { + child = btree_trans_iter_alloc(trans); + child->ip_allocated = ip; + iter->child_idx = child->idx; + + trans->iters_live |= 1ULL << child->idx; + trans->iters_touched |= 1ULL << child->idx; + } + + return child; } static inline void __bch2_trans_iter_free(struct btree_trans *trans, unsigned idx) { + btree_iter_child_free(&trans->iters[idx]); + __bch2_btree_iter_unlock(&trans->iters[idx]); trans->iters_linked &= ~(1ULL << idx); trans->iters_live &= ~(1ULL << idx); trans->iters_touched &= ~(1ULL << idx); - trans->iters_unlink_on_restart &= ~(1ULL << idx); - trans->iters_unlink_on_commit &= ~(1ULL << idx); } -int bch2_trans_iter_free(struct btree_trans *trans, - struct btree_iter *iter) +int bch2_trans_iter_put(struct btree_trans *trans, + struct btree_iter *iter) { - int ret = btree_iter_err(iter); + int ret; - __bch2_trans_iter_free(trans, iter->idx); - return ret; -} + if (IS_ERR_OR_NULL(iter)) + return 0; -int bch2_trans_iter_free_on_commit(struct btree_trans *trans, - struct btree_iter *iter) -{ - int ret = btree_iter_err(iter); + BUG_ON(trans->iters + iter->idx != iter); + BUG_ON(!btree_iter_live(trans, iter)); + + ret = btree_iter_err(iter); - trans->iters_unlink_on_commit |= 1ULL << iter->idx; + if (!(trans->iters_touched & (1ULL << iter->idx)) && + !(iter->flags & BTREE_ITER_KEEP_UNTIL_COMMIT)) + __bch2_trans_iter_free(trans, iter->idx); + + trans->iters_live &= ~(1ULL << iter->idx); return ret; } -static int btree_trans_realloc_iters(struct btree_trans *trans, - unsigned new_size) +int bch2_trans_iter_free(struct btree_trans *trans, + struct btree_iter *iter) { - void *new_iters, *new_updates; + if (IS_ERR_OR_NULL(iter)) + return 0; - BUG_ON(new_size > BTREE_ITER_MAX); + set_btree_iter_dontneed(trans, iter); - if (new_size <= trans->size) - return 0; + return bch2_trans_iter_put(trans, iter); +} - BUG_ON(trans->used_mempool); +noinline __cold +static void btree_trans_iter_alloc_fail(struct btree_trans *trans) +{ - bch2_trans_unlock(trans); + struct btree_iter *iter; + struct btree_insert_entry *i; + char buf[100]; - new_iters = kmalloc(sizeof(struct btree_iter) * new_size + - sizeof(struct btree_insert_entry) * (new_size + 4), - GFP_NOFS); - if (new_iters) - goto success; + trans_for_each_iter(trans, iter) + printk(KERN_ERR "iter: btree %s pos %s%s%s%s %pS\n", + bch2_btree_ids[iter->btree_id], + (bch2_bpos_to_text(&PBUF(buf), iter->pos), buf), + btree_iter_live(trans, iter) ? " live" : "", + (trans->iters_touched & (1ULL << iter->idx)) ? " touched" : "", + iter->flags & BTREE_ITER_KEEP_UNTIL_COMMIT ? " keep" : "", + (void *) iter->ip_allocated); + + trans_for_each_update(trans, i) { + char buf[300]; + + bch2_bkey_val_to_text(&PBUF(buf), trans->c, bkey_i_to_s_c(i->k)); + printk(KERN_ERR "update: btree %s %s\n", + bch2_btree_ids[i->iter->btree_id], buf); + } + panic("trans iter oveflow\n"); +} - new_iters = mempool_alloc(&trans->c->btree_iters_pool, GFP_NOFS); - new_size = BTREE_ITER_MAX; +static struct btree_iter *btree_trans_iter_alloc(struct btree_trans *trans) +{ + struct btree_iter *iter; + unsigned idx; - trans->used_mempool = true; -success: - new_updates = new_iters + sizeof(struct btree_iter) * new_size; + if (unlikely(trans->iters_linked == + ~((~0ULL << 1) << (BTREE_ITER_MAX - 1)))) + btree_trans_iter_alloc_fail(trans); - memcpy(new_iters, trans->iters, - sizeof(struct btree_iter) * trans->nr_iters); - memcpy(new_updates, trans->updates, - sizeof(struct btree_insert_entry) * trans->nr_updates); + idx = __ffs64(~trans->iters_linked); + iter = &trans->iters[idx]; - if (IS_ENABLED(CONFIG_BCACHEFS_DEBUG)) - memset(trans->iters, POISON_FREE, - sizeof(struct btree_iter) * trans->nr_iters + - sizeof(struct btree_insert_entry) * trans->nr_iters); + iter->trans = trans; + iter->idx = idx; + iter->child_idx = U8_MAX; + iter->flags = 0; + iter->nodes_locked = 0; + iter->nodes_intent_locked = 0; + trans->iters_linked |= 1ULL << idx; + return iter; +} - if (trans->iters != trans->iters_onstack) - kfree(trans->iters); +static void btree_iter_copy(struct btree_iter *dst, struct btree_iter *src) +{ + unsigned i; - trans->iters = new_iters; - trans->updates = new_updates; - trans->size = new_size; + __bch2_btree_iter_unlock(dst); + btree_iter_child_free(dst); - if (trans->iters_live) { - trans_restart(); - return -EINTR; - } + memcpy(&dst->flags, &src->flags, + sizeof(struct btree_iter) - offsetof(struct btree_iter, flags)); - return 0; -} + 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)); -void bch2_trans_preload_iters(struct btree_trans *trans) -{ - btree_trans_realloc_iters(trans, BTREE_ITER_MAX); + dst->flags &= ~BTREE_ITER_KEEP_UNTIL_COMMIT; + dst->flags &= ~BTREE_ITER_SET_POS_AFTER_COMMIT; } -static int btree_trans_iter_alloc(struct btree_trans *trans) -{ - unsigned idx = __ffs64(~trans->iters_linked); - - if (idx < trans->nr_iters) - goto got_slot; +struct btree_iter *__bch2_trans_get_iter(struct btree_trans *trans, + unsigned btree_id, struct bpos pos, + unsigned locks_want, + unsigned depth, + unsigned flags) +{ + struct btree_iter *iter, *best = NULL; + struct bpos real_pos, pos_min = POS_MIN; + + if ((flags & BTREE_ITER_TYPE) != BTREE_ITER_NODES && + btree_node_type_is_extents(btree_id) && + !(flags & BTREE_ITER_NOT_EXTENTS) && + !(flags & BTREE_ITER_ALL_SNAPSHOTS)) + flags |= BTREE_ITER_IS_EXTENTS; - if (trans->nr_iters == trans->size) { - int ret = btree_trans_realloc_iters(trans, trans->size * 2); - if (ret) - return ret; - } + if ((flags & BTREE_ITER_TYPE) != BTREE_ITER_NODES && + !btree_type_has_snapshots(btree_id)) + flags &= ~BTREE_ITER_ALL_SNAPSHOTS; - idx = trans->nr_iters++; - BUG_ON(trans->nr_iters > trans->size); + if (!(flags & BTREE_ITER_ALL_SNAPSHOTS)) + pos.snapshot = btree_type_has_snapshots(btree_id) + ? U32_MAX : 0; - trans->iters[idx].idx = idx; -got_slot: - BUG_ON(trans->iters_linked & (1ULL << idx)); - trans->iters_linked |= 1ULL << idx; - return idx; -} + real_pos = pos; -static struct btree_iter *__btree_trans_get_iter(struct btree_trans *trans, - unsigned btree_id, struct bpos pos, - unsigned flags, u64 iter_id) -{ - struct btree_iter *iter; - int idx; + if ((flags & BTREE_ITER_IS_EXTENTS) && + bkey_cmp(pos, POS_MAX)) + real_pos = bpos_nosnap_successor(pos); - BUG_ON(trans->nr_iters > BTREE_ITER_MAX); + trans_for_each_iter(trans, iter) { + if (btree_iter_type(iter) != (flags & BTREE_ITER_TYPE)) + continue; - for (idx = 0; idx < trans->nr_iters; idx++) { - if (!(trans->iters_linked & (1ULL << idx))) + if (iter->btree_id != btree_id) continue; - iter = &trans->iters[idx]; - if (iter_id - ? iter->id == iter_id - : (iter->btree_id == btree_id && - !bkey_cmp(iter->pos, pos))) - goto found; - } - idx = -1; -found: - if (idx < 0) { - idx = btree_trans_iter_alloc(trans); - if (idx < 0) - return ERR_PTR(idx); + if (best) { + int cmp = bkey_cmp(bpos_diff(best->real_pos, real_pos), + bpos_diff(iter->real_pos, real_pos)); - iter = &trans->iters[idx]; - iter->id = iter_id; + if (cmp < 0 || + ((cmp == 0 && btree_iter_keep(trans, iter)))) + continue; + } + + best = iter; + } - bch2_btree_iter_init(trans, iter, btree_id, pos, flags); + if (!best) { + iter = btree_trans_iter_alloc(trans); + bch2_btree_iter_init(trans, iter, btree_id); + } else if (btree_iter_keep(trans, best)) { + iter = btree_trans_iter_alloc(trans); + btree_iter_copy(iter, best); } else { - iter = &trans->iters[idx]; + iter = best; + } - iter->flags &= ~(BTREE_ITER_INTENT|BTREE_ITER_PREFETCH); - iter->flags |= flags & (BTREE_ITER_INTENT|BTREE_ITER_PREFETCH); + trans->iters_live |= 1ULL << iter->idx; + trans->iters_touched |= 1ULL << iter->idx; + + iter->flags = flags; + + iter->snapshot = pos.snapshot; + + /* + * If the iterator has locks_want greater than requested, we explicitly + * do not downgrade it here - on transaction restart because btree node + * split needs to upgrade locks, we might be putting/getting the + * iterator again. Downgrading iterators only happens via an explicit + * bch2_trans_downgrade(). + */ + + locks_want = min(locks_want, BTREE_MAX_DEPTH); + if (locks_want > iter->locks_want) { + iter->locks_want = locks_want; + btree_iter_get_locks(iter, true, _THIS_IP_); } - BUG_ON(iter->btree_id != btree_id); - BUG_ON(trans->iters_live & (1ULL << idx)); - trans->iters_live |= 1ULL << idx; - trans->iters_touched |= 1ULL << idx; + while (iter->level != depth) { + btree_node_unlock(iter, iter->level); + iter->l[iter->level].b = BTREE_ITER_NO_NODE_INIT; + iter->uptodate = BTREE_ITER_NEED_TRAVERSE; + if (iter->level < depth) + iter->level++; + else + iter->level--; + } - BUG_ON(iter->btree_id != btree_id); - BUG_ON((iter->flags ^ flags) & BTREE_ITER_TYPE); + iter->min_depth = depth; - return iter; -} + bch2_btree_iter_set_pos(iter, pos); + btree_iter_set_search_pos(iter, real_pos); -struct btree_iter *__bch2_trans_get_iter(struct btree_trans *trans, - enum btree_id btree_id, - struct bpos pos, unsigned flags, - u64 iter_id) -{ - struct btree_iter *iter = - __btree_trans_get_iter(trans, btree_id, pos, flags, iter_id); + trace_trans_get_iter(_RET_IP_, trans->ip, + btree_id, + &real_pos, locks_want, iter->uptodate, + best ? &best->real_pos : &pos_min, + best ? best->locks_want : U8_MAX, + best ? best->uptodate : U8_MAX); - if (!IS_ERR(iter)) - bch2_btree_iter_set_pos(iter, pos); return iter; } @@ -1798,60 +2206,58 @@ struct btree_iter *bch2_trans_get_node_iter(struct btree_trans *trans, unsigned flags) { struct btree_iter *iter = - __btree_trans_get_iter(trans, btree_id, pos, - flags|BTREE_ITER_NODES, 0); - unsigned i; + __bch2_trans_get_iter(trans, btree_id, pos, + locks_want, depth, + BTREE_ITER_NODES| + BTREE_ITER_NOT_EXTENTS| + BTREE_ITER_ALL_SNAPSHOTS| + flags); - BUG_ON(IS_ERR(iter)); BUG_ON(bkey_cmp(iter->pos, pos)); - - iter->locks_want = locks_want; - iter->level = depth; - - for (i = 0; i < ARRAY_SIZE(iter->l); i++) - iter->l[i].b = NULL; - iter->l[iter->level].b = BTREE_ITER_NOT_END; + BUG_ON(iter->locks_want != min(locks_want, BTREE_MAX_DEPTH)); + BUG_ON(iter->level != depth); + BUG_ON(iter->min_depth != depth); + iter->ip_allocated = _RET_IP_; return iter; } -struct btree_iter *bch2_trans_copy_iter(struct btree_trans *trans, +struct btree_iter *__bch2_trans_copy_iter(struct btree_trans *trans, struct btree_iter *src) { struct btree_iter *iter; - int i, idx; - - idx = btree_trans_iter_alloc(trans); - if (idx < 0) - return ERR_PTR(idx); - trans->iters_live |= 1ULL << idx; - trans->iters_touched |= 1ULL << idx; - trans->iters_unlink_on_restart |= 1ULL << idx; + iter = btree_trans_iter_alloc(trans); + btree_iter_copy(iter, src); - iter = &trans->iters[idx]; - - memcpy(&iter->trans, - &src->trans, - (void *) &iter[1] - (void *) &iter->trans); - - for (i = 0; i < BTREE_MAX_DEPTH; i++) - if (btree_node_locked(iter, i)) - six_lock_increment(&iter->l[i].b->lock, - __btree_lock_want(iter, i)); + trans->iters_live |= 1ULL << iter->idx; + /* + * We don't need to preserve this iter since it's cheap to copy it + * again - this will cause trans_iter_put() to free it right away: + */ + set_btree_iter_dontneed(trans, iter); - return &trans->iters[idx]; + return iter; } -void *bch2_trans_kmalloc(struct btree_trans *trans, - size_t size) +void *bch2_trans_kmalloc(struct btree_trans *trans, size_t size) { - void *ret; + size_t new_top = trans->mem_top + size; + void *p; - if (trans->mem_top + size > trans->mem_bytes) { + if (new_top > trans->mem_bytes) { size_t old_bytes = trans->mem_bytes; - size_t new_bytes = roundup_pow_of_two(trans->mem_top + size); - void *new_mem = krealloc(trans->mem, new_bytes, GFP_NOFS); + size_t new_bytes = roundup_pow_of_two(new_top); + void *new_mem; + + WARN_ON_ONCE(new_bytes > BTREE_TRANS_MEM_MAX); + + new_mem = krealloc(trans->mem, new_bytes, GFP_NOFS); + if (!new_mem && new_bytes <= BTREE_TRANS_MEM_MAX) { + new_mem = mempool_alloc(&trans->c->btree_trans_mem_pool, GFP_KERNEL); + new_bytes = BTREE_TRANS_MEM_MAX; + kfree(trans->mem); + } if (!new_mem) return ERR_PTR(-ENOMEM); @@ -1860,38 +2266,22 @@ void *bch2_trans_kmalloc(struct btree_trans *trans, trans->mem_bytes = new_bytes; if (old_bytes) { - trans_restart(); + trace_trans_restart_mem_realloced(trans->ip, _RET_IP_, new_bytes); return ERR_PTR(-EINTR); } } - ret = trans->mem + trans->mem_top; + p = trans->mem + trans->mem_top; trans->mem_top += size; - return ret; -} - -int bch2_trans_unlock(struct btree_trans *trans) -{ - u64 iters = trans->iters_linked; - int ret = 0; - - while (iters) { - unsigned idx = __ffs64(iters); - struct btree_iter *iter = &trans->iters[idx]; - - ret = ret ?: btree_iter_err(iter); - - __bch2_btree_iter_unlock(iter); - iters ^= 1ULL << idx; - } - - return ret; + memset(p, 0, size); + return p; } -inline void bch2_trans_unlink_iters(struct btree_trans *trans, u64 iters) +inline void bch2_trans_unlink_iters(struct btree_trans *trans) { - iters &= trans->iters_linked; - iters &= ~trans->iters_live; + u64 iters = trans->iters_linked & + ~trans->iters_touched & + ~trans->iters_live; while (iters) { unsigned idx = __ffs64(iters); @@ -1901,58 +2291,263 @@ inline void bch2_trans_unlink_iters(struct btree_trans *trans, u64 iters) } } -void __bch2_trans_begin(struct btree_trans *trans) +void bch2_trans_reset(struct btree_trans *trans, unsigned flags) { - u64 iters_to_unlink; - - /* - * On transaction restart, the transaction isn't required to allocate - * all the same iterators it on the last iteration: - * - * Unlink any iterators it didn't use this iteration, assuming it got - * further (allocated an iter with a higher idx) than where the iter - * was originally allocated: - */ - iters_to_unlink = ~trans->iters_live & - ((1ULL << fls64(trans->iters_live)) - 1); + struct btree_iter *iter; - iters_to_unlink |= trans->iters_unlink_on_restart; - iters_to_unlink |= trans->iters_unlink_on_commit; + trans_for_each_iter(trans, iter) { + iter->flags &= ~(BTREE_ITER_KEEP_UNTIL_COMMIT| + BTREE_ITER_SET_POS_AFTER_COMMIT); + iter->should_be_locked = false; + } - trans->iters_live = 0; + bch2_trans_unlink_iters(trans); - bch2_trans_unlink_iters(trans, iters_to_unlink); + trans->iters_touched &= trans->iters_live; - trans->iters_touched = 0; - trans->iters_unlink_on_restart = 0; - trans->iters_unlink_on_commit = 0; + trans->extra_journal_res = 0; trans->nr_updates = 0; trans->mem_top = 0; - bch2_btree_iter_traverse_all(trans); + 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); + } + + if (!(flags & TRANS_RESET_NOUNLOCK)) + bch2_trans_cond_resched(trans); + + if (!(flags & TRANS_RESET_NOTRAVERSE) && + trans->iters_linked) + bch2_btree_iter_traverse_all(trans); } -void bch2_trans_init(struct btree_trans *trans, struct bch_fs *c) +static void bch2_trans_alloc_iters(struct btree_trans *trans, struct bch_fs *c) { - memset(trans, 0, offsetof(struct btree_trans, iters_onstack)); + size_t iters_bytes = sizeof(struct btree_iter) * BTREE_ITER_MAX; + size_t updates_bytes = sizeof(struct btree_insert_entry) * BTREE_ITER_MAX; + void *p = NULL; + + BUG_ON(trans->used_mempool); + +#ifdef __KERNEL__ + p = this_cpu_xchg(c->btree_iters_bufs->iter, NULL); +#endif + if (!p) + p = mempool_alloc(&trans->c->btree_iters_pool, GFP_NOFS); + trans->iters = p; p += iters_bytes; + trans->updates = p; p += updates_bytes; +} + +void bch2_trans_init(struct btree_trans *trans, struct bch_fs *c, + unsigned expected_nr_iters, + size_t expected_mem_bytes) + __acquires(&c->btree_trans_barrier) +{ + memset(trans, 0, sizeof(*trans)); trans->c = c; - trans->size = ARRAY_SIZE(trans->iters_onstack); - trans->iters = trans->iters_onstack; - trans->updates = trans->updates_onstack; + trans->ip = _RET_IP_; + + /* + * reallocating iterators currently completely breaks + * bch2_trans_iter_put(), we always allocate the max: + */ + bch2_trans_alloc_iters(trans, c); + + if (expected_mem_bytes) { + trans->mem_bytes = roundup_pow_of_two(expected_mem_bytes); + trans->mem = kmalloc(trans->mem_bytes, GFP_KERNEL|__GFP_NOFAIL); + + if (!unlikely(trans->mem)) { + trans->mem = mempool_alloc(&c->btree_trans_mem_pool, GFP_KERNEL); + trans->mem_bytes = BTREE_TRANS_MEM_MAX; + } + } + + trans->srcu_idx = srcu_read_lock(&c->btree_trans_barrier); + +#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 } int bch2_trans_exit(struct btree_trans *trans) + __releases(&c->btree_trans_barrier) { + struct bch_fs *c = trans->c; + bch2_trans_unlock(trans); - kfree(trans->mem); - if (trans->used_mempool) +#ifdef CONFIG_BCACHEFS_DEBUG + if (trans->iters_live) { + struct btree_iter *iter; + + trans_for_each_iter(trans, iter) + btree_iter_child_free(iter); + } + + if (trans->iters_live) { + struct btree_iter *iter; + + bch_err(c, "btree iterators leaked!"); + trans_for_each_iter(trans, iter) + if (btree_iter_live(trans, iter)) + printk(KERN_ERR " btree %s allocated at %pS\n", + bch2_btree_ids[iter->btree_id], + (void *) iter->ip_allocated); + /* Be noisy about this: */ + bch2_fatal_error(c); + } + + mutex_lock(&trans->c->btree_trans_lock); + list_del(&trans->list); + mutex_unlock(&trans->c->btree_trans_lock); +#endif + + srcu_read_unlock(&c->btree_trans_barrier, trans->srcu_idx); + + bch2_journal_preres_put(&trans->c->journal, &trans->journal_preres); + + if (trans->fs_usage_deltas) { + if (trans->fs_usage_deltas->size + sizeof(trans->fs_usage_deltas) == + REPLICAS_DELTA_LIST_MAX) + mempool_free(trans->fs_usage_deltas, + &trans->c->replicas_delta_pool); + else + kfree(trans->fs_usage_deltas); + } + + if (trans->mem_bytes == BTREE_TRANS_MEM_MAX) + mempool_free(trans->mem, &trans->c->btree_trans_mem_pool); + else + kfree(trans->mem); + +#ifdef __KERNEL__ + /* + * Userspace doesn't have a real percpu implementation: + */ + trans->iters = this_cpu_xchg(c->btree_iters_bufs->iter, trans->iters); +#endif + + if (trans->iters) mempool_free(trans->iters, &trans->c->btree_iters_pool); - else if (trans->iters != trans->iters_onstack) - kfree(trans->iters); + trans->mem = (void *) 0x1; trans->iters = (void *) 0x1; return trans->error ? -EIO : 0; } + +static void __maybe_unused +bch2_btree_iter_node_to_text(struct printbuf *out, + struct btree_bkey_cached_common *_b, + enum btree_iter_type type) +{ + pr_buf(out, " l=%u %s:", + _b->level, bch2_btree_ids[_b->btree_id]); + bch2_bpos_to_text(out, btree_node_pos(_b, type)); +} + +#ifdef CONFIG_BCACHEFS_DEBUG +static bool trans_has_btree_nodes_locked(struct btree_trans *trans) +{ + struct btree_iter *iter; + + trans_for_each_iter(trans, iter) + if (btree_iter_type(iter) != BTREE_ITER_CACHED && + iter->nodes_locked) + return true; + return false; +} +#endif + +void bch2_btree_trans_to_text(struct printbuf *out, struct bch_fs *c) +{ +#ifdef CONFIG_BCACHEFS_DEBUG + struct btree_trans *trans; + struct btree_iter *iter; + struct btree *b; + unsigned l; + + mutex_lock(&c->btree_trans_lock); + list_for_each_entry(trans, &c->btree_trans_list, list) { + if (!trans_has_btree_nodes_locked(trans)) + continue; + + pr_buf(out, "%i %ps\n", trans->pid, (void *) trans->ip); + + trans_for_each_iter(trans, iter) { + if (!iter->nodes_locked) + continue; + + pr_buf(out, " iter %u %c %s:", + iter->idx, + btree_iter_type(iter) == BTREE_ITER_CACHED ? 'c' : 'b', + bch2_btree_ids[iter->btree_id]); + bch2_bpos_to_text(out, iter->pos); + pr_buf(out, "\n"); + + for (l = 0; l < BTREE_MAX_DEPTH; l++) { + if (btree_node_locked(iter, l)) { + pr_buf(out, " %s l=%u ", + btree_node_intent_locked(iter, l) ? "i" : "r", l); + bch2_btree_iter_node_to_text(out, + (void *) iter->l[l].b, + btree_iter_type(iter)); + pr_buf(out, "\n"); + } + } + } + + b = READ_ONCE(trans->locking); + if (b) { + iter = &trans->iters[trans->locking_iter_idx]; + pr_buf(out, " locking iter %u %c l=%u %s:", + trans->locking_iter_idx, + btree_iter_type(iter) == BTREE_ITER_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_iter_node_to_text(out, + (void *) b, + btree_iter_type(iter)); + pr_buf(out, "\n"); + } + } + mutex_unlock(&c->btree_trans_lock); +#endif +} + +void bch2_fs_btree_iter_exit(struct bch_fs *c) +{ + mempool_exit(&c->btree_trans_mem_pool); + mempool_exit(&c->btree_iters_pool); + cleanup_srcu_struct(&c->btree_trans_barrier); +} + +int bch2_fs_btree_iter_init(struct bch_fs *c) +{ + unsigned nr = BTREE_ITER_MAX; + + INIT_LIST_HEAD(&c->btree_trans_list); + mutex_init(&c->btree_trans_lock); + + return init_srcu_struct(&c->btree_trans_barrier) ?: + mempool_init_kmalloc_pool(&c->btree_iters_pool, 1, + sizeof(struct btree_iter) * nr + + sizeof(struct btree_insert_entry) * nr) ?: + mempool_init_kmalloc_pool(&c->btree_trans_mem_pool, 1, + BTREE_TRANS_MEM_MAX); +}