Re: [v2 PATCH 3/10] rhashtable: Allow hashfn to be unset

On 03/22/15 at 11:04pm, Herbert Xu wrote:

On Sun, Mar 22, 2015 at 11:55:05AM +0000, Thomas Graf wrote:

quoted

On 03/22/15 at 07:04pm, Herbert Xu wrote:

quoted

@@ -134,6 +136,7 @@ struct rhashtable {
 	struct bucket_table __rcu	*tbl;
 	atomic_t			nelems;
 	bool                            being_destroyed;
+	unsigned int			key_len;

Why is this needed? It looks like you're always initializing this
with ht->p.key_len

It's ht->p.key_len/4 if we use jhash2.

Sure but why not just store key_len/4 in ht->p.key_len then if you
opt in to jhash2() in rhashtable_init()?

quoted

quoted

+	if (!__builtin_constant_p(params.key_len))
+		hash = ht->p.hashfn(key, ht->key_len, tbl->hash_rnd);

I don't understand this. It looks like you only consider
params->key_len if it's constant.

If params->key_len is not constant, then params == ht->p.

I must be missing something obvious. Who guarantees that? I can see
that's true for the current callers but what prevents anybody from
using rhashtable_lookup_fast() with a key length not known at compile
time and pass it as rhashtable_params?

quoted

quoted

+	else if (params.key_len) {
+		unsigned key_len = params.key_len;
+
+		if (params.hashfn)
+			hash = params.hashfn(key, key_len, tbl->hash_rnd);
+		else if (key_len & (sizeof(u32) - 1))
+			hash = jhash(key, key_len, tbl->hash_rnd);
+		else
+			hash = jhash2(key, key_len / sizeof(u32),
+				      tbl->hash_rnd);
+	} else {
+		unsigned key_len = ht->p.key_len;
+
+		if (params.hashfn)
+			hash = params.hashfn(key, key_len, tbl->hash_rnd);
+		else
+			hash = jhash(key, key_len, tbl->hash_rnd);

Why don't we opt-in to jhash2 in this case?

Because if key_len == 0 it means that key_len is not known at
compile-time.

I still don't get this. Why do we fall back to jhash2() if
params.key_len is set but not if only ht->p.key_len is set?