Re: [PATCH net-next 2/3] net: netfilter: Add RFC-7597 Section 5.1 PSID support

[Cole Dishington]

Hello,

Thanks for your time reviewing!

On Wed, 2021-08-25 at 19:05 +0200, Pablo Neira Ayuso wrote:
> Hi,
> 
> On Mon, Aug 09, 2021 at 04:10:36PM +1200, Cole Dishington wrote:
> > Adds support for masquerading into a smaller subset of ports -
> > defined by the PSID values from RFC-7597 Section 5.1. This is part of
> > the support for MAP-E and Lightweight 4over6, which allows multiple
> > devices to share an IPv4 address by splitting the L4 port / id into
> > ranges.
> > 
> > Co-developed-by: Anthony Lineham <anthony.lineham@xxxxxxxxxxxxxxxxxxx>
> > Signed-off-by: Anthony Lineham <anthony.lineham@xxxxxxxxxxxxxxxxxxx>
> > Co-developed-by: Scott Parlane <scott.parlane@xxxxxxxxxxxxxxxxxxx>
> > Signed-off-by: Scott Parlane <scott.parlane@xxxxxxxxxxxxxxxxxxx>
> > Signed-off-by: Blair Steven <blair.steven@xxxxxxxxxxxxxxxxxxx>
> > Signed-off-by: Cole Dishington <Cole.Dishington@xxxxxxxxxxxxxxxxxxx>
> > Reviewed-by: Florian Westphal <fw@xxxxxxxxx>
> [...]
> 
> Looking at the userspace logic:
> 
> https://scanmail.trustwave.com/?c=20988&d=6vim4fcVLjPkIbLUDqz3Tj2W4gXWNCkYa5llWggBjA&u=https%3a%2f%2fpatchwork%2eozlabs%2eorg%2fproject%2fnetfilter-devel%2fpatch%2f20210716002219%2e30193-1-Cole%2eDishington%40alliedtelesis%2eco%2enz%2f
> 
> Chunk extracted from void parse_psid(...)
> 
> >        offset = (1 << (16 - offset_len));
> 
> Assuming offset_len = 6, then you skip 0-1023 ports, OK.
> 
> >        psid = psid << (16 - offset_len - psid_len);
> 
> This psid calculation is correct? Maybe:
> 
>         psid = psid << (16 - offset_len);

PSID port numbers have the form
[offset|PSID|j]
and
16 = offset_length + PSID_length + j_length.
The PSID calculation above is bit shifting the passed psid up j_length.

The userspace tool accepts the unshifted psid to be consistent with how RFC7597 specified it (see RFC7597 Appendix A. Examples).

> 
> instead?
> 
>         psid=0  =>      0 << (16 - 6) = 1024
>         psid=1  =>      1 << (16 - 6) = 2048
> 
> This is implicitly assuming that 64 PSIDs are available, each of them
> taking 1024 ports, ie. psid_len is 6 bits. But why are you subtracting
> the psid_len above?
> 
> >        /* Handle the special case of no offset bits (a=0), so offset loops */
> >        min = psid;
> 
> OK, this line above is the minimal port in the range
> 
> >        if (offset)
> >                min += offset;
> 
> ... which is incremented by the offset (to skip the 0-1023 ports).
> 
> >       r->min_proto.all = htons(min);
> >       r->max_proto.all = htons(min + ((1 << (16 - offset_len - psid_len)) - 1));
> 
> Here, you subtract psid_len again, not sure why.

Each PSID port range is made up of many smaller contiguous port sub-ranges  (except for the special case of offset_len = 0) e.g. for PSID=0x34,psid_length=8,psid_offset=6 the ranges are 1232-1235, 2256-2259, ..., 63696-63699, 64720-64723 (Taken from rfc7597 Appendix A. Examples).
The above calculation is selecting the first sub-range. Max is computed by finding j_length and filling it with 1's.

> 
> >       r->base_proto.all = htons(offset);
> 
> base is set to offset, ie. 1024.
> 
> >       r->flags |= NF_NAT_RANGE_PSID;
> >       r->flags |= NF_NAT_RANGE_PROTO_SPECIFIED;
> 
> Now looking at the kernel side.
> 
> > diff --git a/net/netfilter/nf_nat_masquerade.c b/net/netfilter/nf_nat_masquerade.c
> > index 8e8a65d46345..19a4754cda76 100644
> > --- a/net/netfilter/nf_nat_masquerade.c
> > +++ b/net/netfilter/nf_nat_masquerade.c
> > @@ -55,8 +55,31 @@ nf_nat_masquerade_ipv4(struct sk_buff *skb, unsigned int hooknum,
> >  	newrange.flags       = range->flags | NF_NAT_RANGE_MAP_IPS;
> >  	newrange.min_addr.ip = newsrc;
> >  	newrange.max_addr.ip = newsrc;
> > -	newrange.min_proto   = range->min_proto;
> > -	newrange.max_proto   = range->max_proto;
> > +
> > +	if (range->flags & NF_NAT_RANGE_PSID) {
> > +		u16 base = ntohs(range->base_proto.all);
> > +		u16 min =  ntohs(range->min_proto.all);
> > +		u16 off = 0;
> > +
> > +		/* xtables should stop base > 2^15 by enforcement of
> > +		 * 0 <= offset_len < 16 argument, with offset_len=0
> > +		 * as a special case inwhich base=0.
> 
> I don't understand this comment.

This is a sanity check. The userspace tool restricts offset_len to the specified range and since base = 2^(16 - offset_len) (or base = 0 for the special case of offset_len = 16) the below condition should never be true.
However, if base greater than 1<<15 was allowed, a divide by zero error would occur on the block below.

> 
> > +		 */
> > +		if (WARN_ON_ONCE(base > (1 << 15)))
> > +			return NF_DROP;
> > +
> > +		/* If offset=0, port range is in one contiguous block */
> > +		if (base)
> > +			off = prandom_u32_max(((1 << 16) / base) - 1);
> 
> Assuming the example above, base is set to 1024. Then, off is a random
> value between UINT16_MAX (you expressed this as 1 << 16) and the base
> which is 1024 minus 1.
> 
> So this is picking a random off (actually the PSID?) between 0 and 63.
> What about clashes? I mean, two different machines behind the NAT
> might get the same off.
> 
> > +		newrange.min_proto.all   = htons(min + base * off);
> 
> min could be 1024, 2048, 3072... you add base which is 1024 * off.
> 
> Is this duplicated? Both calculated in user and kernel space?

Each PSID value defines many contiguous port sub-ranges. The randomly chosen off selects the ith sub-range for a given PSID e.g. off=1 would select 2256-2259 for rfc7597 Appendix A. Examples.

The userspace tool calculates the min and max of the first sub-range for a given psid, whereas the above randomly selects one of the sub-ranges for a given psid.

j_length determines how large each sub-range will be, so for small j_length values there still is the risk the chosen sub-range will be exhausted.

> 
> > +		newrange.max_proto.all   = htons(ntohs(newrange.min_proto.all) + ntohs(range->max_proto.all) - min);
> 
> I'm stopping here, I'm getting lost.
> 
> My understanding about this RFC is that you would like to split the
> 16-bit ports in ranges to uniquely identify the host behind the NAT.
> 
> Why don't you just you just select the port range from userspace
> utilizing the existing infrastructure? I mean, why do you need this
> kernel patch?

If utilizing existing infrastruture to install PSID port ranges a lot of rules would be required as each PSID port range is made up of many smaller sub-ranges.

e.g. (from rfc7597 Appendix A. Examples)
for psid_length=8,offset_length=6 each PSID would need 63 NF_NAT_RANGE_PROTO_SPECIFIED rules, hence a total of 16128 rules if all the PSIDs were allocated.

> 
> Florian already suggested:
> 
> > Is it really needed to place all of this in the nat core?
> > 
> > The only thing that has to be done in the NAT core, afaics, is to
> > suppress port reallocation attmepts when NF_NAT_RANGE_PSID is set.
> > 
> > Is there a reason why nf_nat_masquerade_ipv4/6 can't be changed instead
> > to do what you want?
> > 
> > AFAICS its enough to set NF_NAT_RANGE_PROTO_SPECIFIED and init the
> > upper/lower boundaries, i.e. change input given to nf_nat_setup_info().
> 
> extracted from:
> 
> https://scanmail.trustwave.com/?c=20988&d=6vim4fcVLjPkIbLUDqz3Tj2W4gXWNCkYa5s0Bg8JjA&u=https%3a%2f%2fpatchwork%2eozlabs%2eorg%2fproject%2fnetfilter-devel%2fpatch%2f20210422023506%2e4651-1-Cole%2eDishington%40alliedtelesis%2eco%2enz%2f