Re: [PATCH v2 net-next 1/2] cls_bpf: introduce integrated actions

[Jamal Hadi Salim <jhs@mojatatu.com>]

Hi Daniel,

On 09/17/15 09:13, Daniel Borkmann wrote:

> Hmm, I don't really agree. With cls_bpf you have non-linear
> classifications as opposed to walking a chain of classifiers:

A chain of classifiers is a better description today (non-linear would
be an appropriate description before cls_bpf ;->).

> worst case, I have to walk through N classifiers just to find
> out that the last one matches that I need to drop - this doesn't
> scale at all.

The scaling reason with that posted example is not
a strong one. You can get good performance with any classifier
for that policy description.
F.E with Alexei's second best classifier:->:

tc filter add dev eth0 parent ffff: protocol arp prio 1 u32\
match all ..
tc filter add dev eth0 parent ffff: protocol ip prio 1 u32\
...

But I do get the gist of your arguement otherwise and some
short circuits are ok as you had earlier.


>Given that we can make this decision right here,
> we can use this fact and have simple return codes provided as
> well.

I think it makes sense for the simple case.
But you have every other opcode in there, not just basic
accept/drop. I am worried this is leading towards an
enclave of bpf do-everything.


cheers,
jamal