bcachefs: Btree key cache coherency

 - Updates to non key cache iterators will now be transparently
   redirected to the key cache for cached btrees.

 - Except when creating new keys: then the update goes to underlying
   btree

For for iterating over a cached btree to work, we need to ensure that if
a key exists in the key cache, it also exists in the btree - otherwise
the iterator code will skip past it and not check the key cache.

Otherwise, for consistency, all updates should go to the same place -
the key cache.
Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>

fs/bcachefs/alloc_background.c

@@ -271,7 +271,8 @@ int bch2_alloc_write(struct btree_trans *trans, struct btree_iter *iter,
 		return PTR_ERR(a);
 
 	bch2_alloc_pack(trans->c, a, *u);
-	return bch2_trans_update(trans, iter, &a->k, trigger_flags);
+	return bch2_trans_update(trans, iter, &a->k, trigger_flags|
+				 BTREE_UPDATE_NO_KEY_CACHE_COHERENCY);
 }
 
 static unsigned bch_alloc_v1_val_u64s(const struct bch_alloc *a)

fs/bcachefs/btree_iter.c

@@ -2188,6 +2188,8 @@ struct bkey_i *__bch2_btree_trans_peek_updates(struct btree_iter *iter)
 			break;
 		if (bpos_cmp(i->k->k.p, iter->path->pos) < 0)
 			continue;
+		if (i->key_cache_already_flushed)
+			continue;
 		if (!ret || bpos_cmp(i->k->k.p, ret->k.p) < 0)
 			ret = i->k;
 	}

fs/bcachefs/btree_key_cache.c

@@ -414,6 +414,7 @@ static int btree_key_cache_flush_pos(struct btree_trans *trans,
 	 * */
 	ret   = bch2_btree_iter_traverse(&b_iter) ?:
 		bch2_trans_update(trans, &b_iter, ck->k,
+				  BTREE_UPDATE_KEY_CACHE_RECLAIM|
 				  BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE|
 				  BTREE_TRIGGER_NORUN) ?:
 		bch2_trans_commit(trans, NULL, NULL,
@@ -555,13 +556,15 @@ bool bch2_btree_insert_key_cached(struct btree_trans *trans,
 	return true;
 }
 
-#ifdef CONFIG_BCACHEFS_DEBUG
-void bch2_btree_key_cache_verify_clean(struct btree_trans *trans,
-			       enum btree_id id, struct bpos pos)
+void bch2_btree_key_cache_drop(struct btree_trans *trans,
+			       struct btree_path *path)
 {
-	BUG_ON(bch2_btree_key_cache_find(trans->c, id, pos));
+	struct bkey_cached *ck = (void *) path->l[0].b;
+
+	ck->valid = false;
+
+	BUG_ON(test_bit(BKEY_CACHED_DIRTY, &ck->flags));
 }
-#endif
 
 static unsigned long bch2_btree_key_cache_scan(struct shrinker *shrink,
 					   struct shrink_control *sc)

fs/bcachefs/btree_key_cache.h

@@ -32,14 +32,8 @@ bool bch2_btree_insert_key_cached(struct btree_trans *,
 			struct btree_path *, struct bkey_i *);
 int bch2_btree_key_cache_flush(struct btree_trans *,
 			       enum btree_id, struct bpos);
-#ifdef CONFIG_BCACHEFS_DEBUG
-void bch2_btree_key_cache_verify_clean(struct btree_trans *,
-				enum btree_id, struct bpos);
-#else
-static inline void
-bch2_btree_key_cache_verify_clean(struct btree_trans *trans,
-				enum btree_id id, struct bpos pos) {}
-#endif
+void bch2_btree_key_cache_drop(struct btree_trans *,
+			       struct btree_path *);
 
 void bch2_fs_btree_key_cache_exit(struct btree_key_cache *);
 void bch2_fs_btree_key_cache_init_early(struct btree_key_cache *);

fs/bcachefs/btree_types.h

@@ -344,6 +344,7 @@ struct btree_insert_entry {
 	bool			cached:1;
 	bool			insert_trigger_run:1;
 	bool			overwrite_trigger_run:1;
+	bool			key_cache_already_flushed:1;
 	/*
 	 * @old_k may be a key from the journal; @old_btree_u64s always refers
 	 * to the size of the key being overwritten in the btree:
@@ -645,6 +646,8 @@ static inline bool btree_type_has_snapshots(enum btree_id id)
 enum btree_update_flags {
 	__BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE,
 	__BTREE_UPDATE_NOJOURNAL,
+	__BTREE_UPDATE_KEY_CACHE_RECLAIM,
+	__BTREE_UPDATE_NO_KEY_CACHE_COHERENCY,
 
 	__BTREE_TRIGGER_NORUN,		/* Don't run triggers at all */
 
@@ -658,6 +661,9 @@ enum btree_update_flags {
 
 #define BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE (1U << __BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE)
 #define BTREE_UPDATE_NOJOURNAL		(1U << __BTREE_UPDATE_NOJOURNAL)
+#define BTREE_UPDATE_KEY_CACHE_RECLAIM	(1U << __BTREE_UPDATE_KEY_CACHE_RECLAIM)
+#define BTREE_UPDATE_NO_KEY_CACHE_COHERENCY	\
+	(1U << __BTREE_UPDATE_NO_KEY_CACHE_COHERENCY)
 
 #define BTREE_TRIGGER_NORUN		(1U << __BTREE_TRIGGER_NORUN)
 

fs/bcachefs/btree_update.h

@@ -76,8 +76,6 @@ int bch2_btree_node_update_key_get_iter(struct btree_trans *,
 int bch2_trans_update_extent(struct btree_trans *, struct btree_iter *,
 			     struct bkey_i *, enum btree_update_flags);
 
-int __must_check bch2_trans_update_by_path(struct btree_trans *, struct btree_path *,
-				   struct bkey_i *, enum btree_update_flags);
 int __must_check bch2_trans_update(struct btree_trans *, struct btree_iter *,
 				   struct bkey_i *, enum btree_update_flags);
 

fs/bcachefs/btree_update_leaf.c

@@ -23,10 +23,15 @@
 #include <linux/prefetch.h>
 #include <linux/sort.h>
 
+static int __must_check
+bch2_trans_update_by_path(struct btree_trans *, struct btree_path *,
+			  struct bkey_i *, enum btree_update_flags);
+
 static inline int btree_insert_entry_cmp(const struct btree_insert_entry *l,
 					 const struct btree_insert_entry *r)
 {
 	return   cmp_int(l->btree_id,	r->btree_id) ?:
+		 cmp_int(l->cached,	r->cached) ?:
 		 -cmp_int(l->level,	r->level) ?:
 		 bpos_cmp(l->k->k.p,	r->k->k.p);
 }
@@ -378,9 +383,14 @@ static inline void do_btree_insert_one(struct btree_trans *trans,
 
 	i->k->k.needs_whiteout = false;
 
-	did_work = !i->cached
-		? btree_insert_key_leaf(trans, i)
-		: bch2_btree_insert_key_cached(trans, i->path, i->k);
+	if (!i->cached)
+		did_work = btree_insert_key_leaf(trans, i);
+	else if (!i->key_cache_already_flushed)
+		did_work = bch2_btree_insert_key_cached(trans, i->path, i->k);
+	else {
+		bch2_btree_key_cache_drop(trans, i->path);
+		did_work = false;
+	}
 	if (!did_work)
 		return;
 
@@ -987,18 +997,6 @@ int __bch2_trans_commit(struct btree_trans *trans)
 			goto out_reset;
 	}
 
-#ifdef CONFIG_BCACHEFS_DEBUG
-	/*
-	 * if BTREE_TRIGGER_NORUN is set, it means we're probably being called
-	 * from the key cache flush code:
-	 */
-	trans_for_each_update(trans, i)
-		if (!i->cached &&
-		    !(i->flags & BTREE_TRIGGER_NORUN))
-			bch2_btree_key_cache_verify_clean(trans,
-					i->btree_id, i->k->k.p);
-#endif
-
 	ret = bch2_trans_commit_run_triggers(trans);
 	if (ret)
 		goto out;
@@ -1369,11 +1367,14 @@ static int need_whiteout_for_snapshot(struct btree_trans *trans,
 	return ret;
 }
 
-int __must_check bch2_trans_update_by_path(struct btree_trans *trans, struct btree_path *path,
-				   struct bkey_i *k, enum btree_update_flags flags)
+static int __must_check
+bch2_trans_update_by_path_trace(struct btree_trans *trans, struct btree_path *path,
+				struct bkey_i *k, enum btree_update_flags flags,
+				unsigned long ip)
 {
 	struct bch_fs *c = trans->c;
 	struct btree_insert_entry *i, n;
+	int ret = 0;
 
 	BUG_ON(!path->should_be_locked);
 
@@ -1388,7 +1389,7 @@ int __must_check bch2_trans_update_by_path(struct btree_trans *trans, struct btr
 		.cached		= path->cached,
 		.path		= path,
 		.k		= k,
-		.ip_allocated	= _RET_IP_,
+		.ip_allocated	= ip,
 	};
 
 #ifdef CONFIG_BCACHEFS_DEBUG
@@ -1409,17 +1410,6 @@ int __must_check bch2_trans_update_by_path(struct btree_trans *trans, struct btr
 	    !btree_insert_entry_cmp(&n, i)) {
 		BUG_ON(i->insert_trigger_run || i->overwrite_trigger_run);
 
-		/*
-		 * This is a hack to ensure that inode creates update the btree,
-		 * not the key cache, which helps with cache coherency issues in
-		 * other areas:
-		 */
-		if (n.cached && !i->cached) {
-			i->k = n.k;
-			i->flags = n.flags;
-			return 0;
-		}
-
 		bch2_path_put(trans, i->path, true);
 		i->flags	= n.flags;
 		i->cached	= n.cached;
@@ -1444,19 +1434,60 @@ int __must_check bch2_trans_update_by_path(struct btree_trans *trans, struct btr
 		}
 	}
 
-	__btree_path_get(n.path, true);
-	return 0;
+	__btree_path_get(i->path, true);
+
+	/*
+	 * If a key is present in the key cache, it must also exist in the
+	 * btree - this is necessary for cache coherency. When iterating over
+	 * a btree that's cached in the key cache, the btree iter code checks
+	 * the key cache - but the key has to exist in the btree for that to
+	 * work:
+	 */
+	if (path->cached &&
+	    bkey_deleted(&i->old_k) &&
+	    !(flags & BTREE_UPDATE_NO_KEY_CACHE_COHERENCY)) {
+		struct btree_path *btree_path;
+
+		i->key_cache_already_flushed = true;
+		i->flags |= BTREE_TRIGGER_NORUN;
+
+		btree_path = bch2_path_get(trans, path->btree_id, path->pos,
+					   1, 0, BTREE_ITER_INTENT);
+
+		ret = bch2_btree_path_traverse(trans, btree_path, 0);
+		if (ret)
+			goto err;
+
+		btree_path->should_be_locked = true;
+		ret = bch2_trans_update_by_path_trace(trans, btree_path, k, flags, ip);
+err:
+		bch2_path_put(trans, btree_path, true);
+	}
+
+	return ret;
+}
+
+static int __must_check
+bch2_trans_update_by_path(struct btree_trans *trans, struct btree_path *path,
+			  struct bkey_i *k, enum btree_update_flags flags)
+{
+	return bch2_trans_update_by_path_trace(trans, path, k, flags, _RET_IP_);
 }
 
 int __must_check bch2_trans_update(struct btree_trans *trans, struct btree_iter *iter,
 				   struct bkey_i *k, enum btree_update_flags flags)
 {
+	struct btree_path *path = iter->update_path ?: iter->path;
+	struct bkey_cached *ck;
+	int ret;
+
 	if (iter->flags & BTREE_ITER_IS_EXTENTS)
 		return bch2_trans_update_extent(trans, iter, k, flags);
 
 	if (bkey_deleted(&k->k) &&
+	    !(flags & BTREE_UPDATE_KEY_CACHE_RECLAIM) &&
 	    (iter->flags & BTREE_ITER_FILTER_SNAPSHOTS)) {
-		int ret = need_whiteout_for_snapshot(trans, iter->btree_id, k->k.p);
+		ret = need_whiteout_for_snapshot(trans, iter->btree_id, k->k.p);
 		if (unlikely(ret < 0))
 			return ret;
 
@@ -1464,8 +1495,45 @@ int __must_check bch2_trans_update(struct btree_trans *trans, struct btree_iter
 			k->k.type = KEY_TYPE_whiteout;
 	}
 
-	return bch2_trans_update_by_path(trans, iter->update_path ?: iter->path,
-					 k, flags);
+	/*
+	 * Ensure that updates to cached btrees go to the key cache:
+	 */
+	if (!(flags & BTREE_UPDATE_KEY_CACHE_RECLAIM) &&
+	    !path->cached &&
+	    !path->level &&
+	    btree_id_cached(trans->c, path->btree_id)) {
+		if (!iter->key_cache_path ||
+		    !iter->key_cache_path->should_be_locked ||
+		    bpos_cmp(iter->key_cache_path->pos, k->k.p)) {
+			if (!iter->key_cache_path)
+				iter->key_cache_path =
+					bch2_path_get(trans, path->btree_id, path->pos, 1, 0,
+						      BTREE_ITER_INTENT|BTREE_ITER_CACHED);
+
+			iter->key_cache_path =
+				bch2_btree_path_set_pos(trans, iter->key_cache_path, path->pos,
+							iter->flags & BTREE_ITER_INTENT);
+
+			ret = bch2_btree_path_traverse(trans, iter->key_cache_path,
+						       BTREE_ITER_CACHED);
+			if (unlikely(ret))
+				return ret;
+
+			ck = (void *) iter->key_cache_path->l[0].b;
+
+			if (test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
+				trace_trans_restart_key_cache_raced(trans->fn, _RET_IP_);
+				btree_trans_restart(trans);
+				return -EINTR;
+			}
+
+			iter->key_cache_path->should_be_locked = true;
+		}
+
+		path = iter->key_cache_path;
+	}
+
+	return bch2_trans_update_by_path(trans, path, k, flags);
 }
 
 void bch2_trans_commit_hook(struct btree_trans *trans,

fs/bcachefs/trace.h

@@ -658,6 +658,12 @@ DEFINE_EVENT(transaction_restart,	trans_restart_mark_replicas,
 	TP_ARGS(trans_fn, caller_ip)
 );
 
+DEFINE_EVENT(transaction_restart,	trans_restart_key_cache_raced,
+	TP_PROTO(const char *trans_fn,
+		 unsigned long caller_ip),
+	TP_ARGS(trans_fn, caller_ip)
+);
+
 DECLARE_EVENT_CLASS(transaction_restart_iter,
 	TP_PROTO(const char *trans_fn,
 		 unsigned long caller_ip,