I followed the whole discussion about vlan filtering broken in the 3.x kernel as on observer (I never had time to fully read all the emails and understand all the details).
In general my gut feeling (and I might be totally wrong) is that packet capture in the linux kernel is becoming more and more broken or at least difficult to support. The whole idea is being able to capture the packets as they are seen or transmitted by the NIC card. Vlan tag is stripped, so the capture piece in the kernel receives the vlan tag as oob information. How does it work for Q-in-Q? And as you said, how does it work with the (already broken) vlan support in BPF compiler. This seems like the perfect recipe for more issues in the future. My 2 cents. GV -----Original Message----- From: paul.j.pea...@gmail.com [mailto:paul.j.pea...@gmail.com] On Behalf Of Paul Pearce Sent: Friday, February 01, 2013 12:51 PM To: Gianluca Varenni Cc: Bill Fenner; Michael Richardson; tcpdump-workers@lists.tcpdump.org; Francesco Ruggeri Subject: Re: [tcpdump-workers] "not vlan" filter expression broken catastrophically! I'd like to point out that vlan filtering in general is completely broken under Linux 3x (as discussed several times on this list). In Linux 3x they began stripping the vlan headers off of RX packets and setting BPF ancillary flags, but not doing the same on TX packets. Since the vlan tags are missing when RX packets reach the kernel filter it means that stock libpcap plus any linux 3x kernel can only see TX vlan tagged packets. A recent (3.8 I believe) patch added the ability to use BPF to poke at the vlan ancillary fields, and Ani RFC'd a patch to on this list to shift vlan filtering to using the ancillary fields rather than offsetting into the header. But even with that patch since RX and TX paths are different, it's still not fixed. You could imagine extending Ani's patch to check for the vlan ancillary fields and if not set then look at the headers, but that would mean the filter: vlan X or vlan Y would have different behavior on RX vs TX packets because of the pointer into the header advancing when it encounters a vlan tag on TX, but not RX. In my humble (uneducated) opinion the correct fix is to get linux to move to setting the vlan ancillary fields on TX packets as they do now on RX packets, which would simplify things a lot for libpcap. But that idea got a lot of pushback on the net-dev list. I didn't fully understand their distinction as to why it was ok on RX vs TX, and they never answered when I asked. -Paul On Fri, Feb 1, 2013 at 8:51 AM, Gianluca Varenni <gianluca.vare...@riverbed.com> wrote: > The problem is that if you change the behavior of the vlan keyword, > you potentially break a lot of applications that are based on the old > buggy behavior :-( > > -----Original Message----- > From: fen...@gmail.com [mailto:fen...@gmail.com] On Behalf Of Bill > Fenner > Sent: Friday, February 01, 2013 4:49 AM > To: Gianluca Varenni > Cc: Ani Sinha; tcpdump-workers@lists.tcpdump.org; Michael Richardson; > Francesco Ruggeri > Subject: Re: [tcpdump-workers] "not vlan" filter expression broken > catastrophically! > > On Thu, Jan 31, 2013 at 7:20 PM, Gianluca Varenni > <gianluca.vare...@riverbed.com> wrote: >> To be totally honest, I think the whole way in which vlans are >> managed in the filters is quite nonsense. The underlying problem is >> that normally a BPF filter is an "or" or "and" combination of >> disjoint filters, so if I write "filterA" or "filterB" I assume that >> the two filters are disjoints, so >> >> "filterA or filterB" should be equivalent to "filterB or filterA" >> >> This is not true when using the "vlan" keyword. Vlan sticks globally and >> increments the offset of the L3 header unconditionally of two bytes, no >> turning back. >> >> For example "ip or vlan 14" is different than "vlan 14 or ip" > > We have wanted to fix the vlan support ever since it was added. If I > remember right we even talked about not adding it and waiting to do it right. > It's definitely a hack, the vlan offset info should be associative and only > apply to anything that is "and"ed with the vlan keyword. Sadly, the current > structure of the parser / code generator do not lend themselves to that. > > The global nature of the vlan offset is something that nobody is happy with. > All it will take to fix it is to rewrite the grammar parser and filter > generation code. > > Bill > > >> -----Original Message----- >> From: tcpdump-workers-boun...@lists.tcpdump.org >> [mailto:tcpdump-workers-boun...@lists.tcpdump.org] On Behalf Of Ani >> Sinha >> Sent: Thursday, January 31, 2013 3:42 PM >> To: tcpdump-workers@lists.tcpdump.org >> Cc: Bill Fenner; Michael Richardson; Francesco Ruggeri >> Subject: [tcpdump-workers] "not vlan" filter expression broken >> catastrophically! >> >> hello folks : >> >> As you guys have been aware, I am hacking libpcap for a while. Me and Bill >> noticed something seriously broken for any filter expression that has a "not >> vlan" in it. For example, take a look at the filter code generated by >> libpcap with an expression like "not vlan and tcp port 80" : >> >> BpfExpression '(not vlan and tcp port 80)' >> { 0x28, 0, 0, 0x0000000c }, //(000) ldh [12] >> { 0x15, 19, 0, 0x00008100 }, //(001) jeq #0x8100 jt 21 jf 2 >> { 0x28, 0, 0, 0x00000010 }, //(002) ldh [16] >> { 0x15, 0, 6, 0x000086dd }, //(003) jeq #0x86dd jt 4 jf 10 >> { 0x30, 0, 0, 0x00000018 }, //(004) ldb [24] >> { 0x15, 0, 15, 0x00000006 }, //(005) jeq #0x6 jt 6 jf 21 >> { 0x28, 0, 0, 0x0000003a }, //(006) ldh [58] >> { 0x15, 12, 0, 0x00000050 }, //(007) jeq #0x50 jt 20 jf 8 >> { 0x28, 0, 0, 0x0000003c }, //(008) ldh [60] >> { 0x15, 10, 11, 0x00000050 }, //(009) jeq #0x50 jt 20 jf 21 >> { 0x15, 0, 10, 0x00000800 }, //(010) jeq #0x800 jt 11 jf 21 >> { 0x30, 0, 0, 0x0000001b }, //(011) ldb [27] >> { 0x15, 0, 8, 0x00000006 }, //(012) jeq #0x6 jt 13 jf 21 >> { 0x28, 0, 0, 0x00000018 }, //(013) ldh [24] >> { 0x45, 6, 0, 0x00001fff }, //(014) jset #0x1fff jt 21 jf 15 >> { 0xb1, 0, 0, 0x00000012 }, //(015) ldxb 4*([18]&0xf) >> { 0x48, 0, 0, 0x00000012 }, //(016) ldh [x + 18] >> { 0x15, 2, 0, 0x00000050 }, //(017) jeq #0x50 jt 20 jf 18 >> { 0x48, 0, 0, 0x00000014 }, //(018) ldh [x + 20] >> { 0x15, 0, 1, 0x00000050 }, //(019) jeq #0x50 jt 20 jf 21 >> { 0x6, 0, 0, 0x0000ffff }, //(020) ret #65535 >> { 0x6, 0, 0, 0x00000000 }, //(021) ret #0 >> >> >> As you can see, it loads offset 12 (ethertype). For vlan packets, it jumps >> to #21 and returns false right away. However, for packets that are not vlan >> tagged, it goes to #2 which loads offset 16 in the packet. Notice that this >> is wrong! The offsets should be incremented by 4 only for vlan tagged >> packets and not for non-vlan packets. The problem is that in gencode.c, the >> off_linktype increments by 4 unconditionally whether or not the packet >> actually contains a vlan tag. We do not want to increment this offset if >> "not vlan" is true. So the above filter code is generated wrong. >> >> I just wanted to point this out to folks who wishes to dig in and fix it. I >> do not have time right now to think of a proper solution. It would seem >> using unconditional increments of offsets like off_linktype below the parser >> is not going to work. How do you know if the parser is going to take your >> code generated from the "vlan" expression and just negate it? Or may be we >> can hack another rule in grammar.y. I don't know. >> >> cheers, >> ani >> _______________________________________________ >> tcpdump-workers mailing list >> tcpdump-workers@lists.tcpdump.org >> https://lists.sandelman.ca/mailman/listinfo/tcpdump-workers >> _______________________________________________ >> tcpdump-workers mailing list >> tcpdump-workers@lists.tcpdump.org >> https://lists.sandelman.ca/mailman/listinfo/tcpdump-workers > _______________________________________________ > tcpdump-workers mailing list > tcpdump-workers@lists.tcpdump.org > https://lists.sandelman.ca/mailman/listinfo/tcpdump-workers _______________________________________________ tcpdump-workers mailing list tcpdump-workers@lists.tcpdump.org https://lists.sandelman.ca/mailman/listinfo/tcpdump-workers
