Re: [PATCH nft 2/3] meta: add prandom matching

[Florian Westphal <fw@strlen.de>]

Pablo Neira Ayuso <pablo@netfilter.org> wrote:
> On Mon, Feb 15, 2016 at 01:54:46PM +0100, Florian Westphal wrote:
> > Patrick McHardy <kaber@trash.net> wrote:
> > 
> > Hi Patrick
> > 
> > > On 04.02, Florian Westphal wrote:
> > > > In fact, doing the scaling via precision_type seems to
> > > > be a lot simpler as then its applied only in this one case of the
> > > > prandom META_TEMPLATE while keeping this detail limited to meta.c.
> > > 
> > > Yes, on second thought I agree, sorry. Maybe the work is not lost though,
> > > what does seem to make sense is to use a float basetype and derive your
> > > probability type from that.
> > 
> > I can do this.
> > However, I don't currently see any other type that could be derived from
> > that.  Would you be OK with leaving things as-is and adding a float
> > type later on once a use case presents itself?
> 
> I also think you can add a new TYPE_FLOAT.

Yes, but what I'm asking is: 'What for'?

> Then, from the evaluation step make sure that META_PRANDOM is under
> the valid limits (0, 1].

Thats not so simple.  META_PRANDOM is scaled so its represented
as 1.0 == UINT32_MAX and 0 = 0.

We can't do that for TYPE_FLOAT since it means that it could not
represent values > 1.0 .

Doing the scaling in the eval step is possible but its a bit ugly.

> These TYPE_* will be part of the public API of the high level library
> at some point, they describe the datatype that are used in set
> definitions in the kernel (through the NFTA_SET_DATATYPE netlink
> attribute and the new NFTA_SET_USERDATA through TLVs).

I understand, but,  the proposed float and probability types
are very different, and allow for very little re-use.

For example on printing the probability type has to undo the scaling
so we print 1.0 instead of $bignum.