On Thu, Aug 13, 2020 at 5:52 AM Jamal Hadi Salim <j...@mojatatu.com> wrote:
>
> The _main_ requirement is to scale to a large number of filters
> (a million is a good handwave number). Scale means
> 1) fast datapath lookup time + 2) fast insertion/deletion/get/dump
> from control/user space.
> fwmark is good at all these goals today for #2. It is good for #1 for
> maybe 1K rules (limitation is the 256 buckets, constrained by rcu
> trickery). Then you start having collisions in a bucket and your
> lookup requires long linked list walks.
>
> Generally something like a hash table with sufficient number of buckets
> will work out ok.
> There maybe other approaches (IDR in the kernel looks interesting,
> but i didnt look closely).
>
> So to the implementation issue:
> Major issue is removing ambiguity while at the same time trying
> to get good performance.
>
> Lets say we decided to classify skbmark and skbhash at this point.
> For a hash table, one simple approach is to set
> lookupkey = hash<<32|mark
>
> the key is used as input to the hash algo to find the bucket.
>
> There are two outstanding challenges in my mind:
>
> 1) To use the policy like you describe above
> as an example:
>
> $TC filter add dev $DEV1 parent ffff: protocol ip prio 3 handle 1
> skb hash Y flowid 1:12 action ok
>
> and say you receive a packet with both skb->hash and skb->mark set
> Then there is ambiguity
>
> How do you know whether to use hash or mark or both
> for that specific key?
Hmm, you can just unconditionally pass skb->hash and skb->mark,
no? Something like:
if (filter_parameter_has_hash) {
match skb->hash with cls->param_hash
}
if (filter_parameter_has_mark) {
match skb->mark with cls->param_mark
}
fw_classify() uses skb->mark unconditionally anyway, without checking
whether it is set or not first.
But if filters were put in a global hashtable, the above would be
much harder to implement.
> You can probably do some trick but I cant think of a cheap way to
> achieve this goal. Of course this issue doesnt exist if you have
> separate classifiers.
>
> 2) If you decide tomorrow to add tcindex/prio etc, you will have to
> rework this as well.
>
> #2 is not as a big deal as #1.
Well, I think #2 is more serious than #1, if we have to use a hashtable.
(If we don't have to, then it would be much easier to extend, of course.)
THanks.
