On 9/2/25 3:18 PM, Florian Westphal wrote: > Eric Woudstra <ericwouds@xxxxxxxxx> wrote: >>> Thats because of implicit dependency insertion on userspace side: >>> # ip saddr 1.2.3.4 counter ip daddr 3.4.5.6 >>> bridge test-bridge input >>> [ meta load protocol => reg 1 ] >>> [ cmp eq reg 1 0x00000008 ] >>> [ payload load 4b @ network header + 12 => reg 1 ] >>> ... >>> >>> So, if userspace would NOT do that it would 'just work'. >>> >>> Pablo, whats your take on this? >>> We currently don't have a 'nhproto' field in nft_pktinfo >>> and there is no space to add one. >>> >>> We could say that things work as expected, and that >>> ip saddr 1.2.3.4 >>> >>> should not magically match packets in e.g. pppoe encap. > > FTR, I think 'ip saddr 1.2.3.4' (standalone with no other info), > should NOT match inside a random l2 tunnel. > >>> I suspect it will start to work if you force it to match in pppoe, e.g. >>> ether type 0x8864 ip saddr ... >>> >>> so nft won't silently add the skb->protocol dependency. >>> >>> Its not a technical issue but about how matching is supposed to work >>> in a bridge. >>> >>> If its supposed to work automatically we need to either: >>> 1. munge skb->protocol in kernel, even tough its wrong (we don't strip >>> the l2 headers). >>> 2. record the real l3 protocol somewhere and make it accessible, then >>> fix the dependency generation in userspace to use the 'new way' (meta >>> l3proto)? >>> 3. change the dependency generation to something else. >>> But what? 'ether type ip' won't work either for 8021ad etc. >>> 'ip version' can't be used for arp. >>> >> >> Hi Florian, >> >> Did you get any information on how to handle this issue? > > Did you check if you can get it to match if you add the relevant > l3 dependency in the rule? > > I don't think we should (or can) change how the rules get evaluated by > making 'ip saddr' match on other l2 tunnel protocols by default. > > It is even incompatible with any exiting rulesets, consider e.g. > "ip daddr 1.2.3.4 drop" on a bridge, now this address becomes > unreachable but it works before your patch (if the address is found in > e.g. pppoe header). > > 'ip/ip6' should work as expected as long as userspace provides > the correct ether type and dependencies. > > I.e., what this patch adds as C code should work if being provided > as part of the rule. > > What might make sense is to add the ppp(oe) header to src/proto.c > in nftables so users that want to match the header following ppp > one don't have to use raw payload match syntax. > > What might also make sense is to either add a way to force a call > to nft_set_pktinfo_ipv4_validate() from the ruleset, or take your > patch but WITHOUT the skb->protocol munging. > > However, due to the number of possible l2 header chain combinations > I'm not sure we should bother with trying to add all of them. > > I worry we would end up turning nft_do_chain_bridge() preamble or > nft_set_pktinfo() into some kind of l2 packet dissector. > > Maybe one way forward is to introduce > > NFT_META_BRI_INET_VALIDATE > > nft add rule ... meta inet validate ... > (just an idea, come up with better names...) > > We'd have to add NFT_PKTINFO_L3PROTO flag to > include/net/netfilter/nf_tables.h. > (or, alternatively NFT_PKTINFO_UNSPEC). > > Then, set this flag in struct nft_pktinfo, from > nft_set_pktinfo_ipv4|6_validate (or from nft_set_pktinfo_unspec). > > NFT_META_BRI_INET_VALIDATE, would call nft_set_pktinfo_ipv4_validate > or nft_set_pktinfo_ipv6_validate depending on iph->version and set > NFT_BREAK verdict if the flag is still absent. > > **USERSPACE IS RESPONSIBLE** to prevent arp packets from entering > this expression. If they do, then header validation should fail > but there would be an off-chance that the garbage is also a valid > ipv4 or ipv6 packet. I'll try to recap what this would mean for this patch, correct me if I am wrong. 'ip saddr 1.2.3.4' should NOT match inside a random l2 tunnel. That makes the 'issue' expected behavior. If the user must need to match ip(v6) address, it must be solved from userspace, modifying the rule. This patch (3/3) can be applied as-is, with only one change: remove munging skb->protocol.
