Implementation of Sampling for BPF

[Vadim Goncharov]

07.01.08 @ 04:01 Peter Wood wrote:

>> I don't think that modifying bpf.c is good solution, as userland is not  
>> the only consumer of BPF, think, for example, about ng_bpf. Moreover,  
>> what is the purpose of sampling, after all? BPF was never intended to  
>> be reliable every-packet solution.
>
> Certainly other things do use BPF, however in my case I'm not using  
> them, and in the 1 in X solution I have developed so far it can be  
> turned on and off and if it's of huge concern could be put between  
> defines and a kernel config option be required to include it.

It's the question of doing things correctly(tm) so they are appropriate  
for inclusion into the main src tree of the FreeBSD Project - this must be  
universal enough to meet other people needs and to be supported. You of  
course are free to do any patches at your locals site for your individual  
needs - many people do that customization on their own.

> I'm not looking to transform BPF into a solution to reliably sample  
> every packet, I am looking at attempting to define which packets it  
> discards so that there is an equal chance of sampling something that  
> happens, rather then an unknown/unpredictable chance.

So what if a malicious packet will be skipped due sampling, packet which  
is by other means undistinguishable from others before detailed analysis?

> I wanted to stop the packet being sent to BPF as high up the kernel  
> chain as possible as to save as much CPU time as possible. There's no  
> point in capturing everything we can and then having the user land  
> program selectively chuck stuff when it could be done before all the  
> various copying/switching/etc.

Low in chain instead of high, you mean? That's of course no point to sort  
out things in userland, but that's properties of given BPF program to  
filter - how much the userland program wants to receive before detailed  
analysis.

>> If you are monitoring in userland, Snort of course will not have enough  
>> time to process all of your data, so why not simply put at least two  
>> machines in parallel, one for each mirrored line?
>
> 1) This doesn't scale, in the next six to twelve months I'm going to be  
> presented with a 10Gb uplink to our regional network. Now I know I'm  
> going to have issues when that link reaches ~40% capacity anyway, but  
> one thing at a time.
>
> 2) We don't have the machine room heat or power capacity spare to run  
> more servers, and there are other projects that require capacity that  
> are in the waiting list way ahead of mine.

> I'd love to buy a commercial hardware solution, unfortunately my budget  
> is short by about $750k. So here I am with my favourite OS instead. God  
> knows I've benefited from using FreeBSD, as has the institute I work  
> for, at least if I do it properly I can say "guys, it's yours if you  
> want it".

Putting as many servers as needed does scale well if you need only sampled  
data - just put an appropriate sampler/load balancer before them. And  
using FreeBSD on that servers will be cheaper than commercial hardware  
solution, too.

> 3) Because of our constraints we are satisfied with sampled data, we  
> don't need full streams, but we would like controlled sampled data.

Why sample is enough to you? What exactly do you need? May be you'd rather  
write some simpler expressions for in-kernel filtering instead of  
heavy-weighted Snort?

-- 
WBR, Vadim Goncharov