On 05/22/2019 01:52 AM, Matthew Cover wrote:
> > > __sk_buff has a member tc_classid which I'm interested in accessing from
> > > the skb bpf context.
> > >
> > > A bpf program which accesses skb->tc_classid compiles, but fails
> > > verification; the specific failure is "invalid bpf_context access".
> > >
> > > if (skb->tc_classid != 0)
> > > return 1;
> > > return 0;
> > >
> > > Some of the tests in tools/testing/selftests/bpf/verifier/ (those on
> > > tc_classid) further confirm that this is, in all likelihood, intentional
> > > behavior.
> > >
> > > The very similar bpf program which instead accesses skb->mark works as
> > > desired.
> > >
> > > if (skb->mark != 0)
> > > return 1;
> > > return 0;
> >
> > You should be able to access skb->tc_classid, perhaps you're using the
> > wrong program
> > type? BPF_PROG_TYPE_SCHED_CLS is supposed to work (if not we'd have a
> > regression).
> >
>
> I am in fact using BPF_PROG_TYPE_SOCKET_FILTER and using the program as
> PACKET_FANOUT_DATA with PACKET_FANOUT_EBPF.
>
> I have been working on a series of utils which leverage PACKET_FANOUT to
> provide various per-socket-fd (per-cpu, per-queue,
> per-rx-flow-hash-indirection-table-idx) statistics and pcap files. While
> playing with PACKET_FANOUT_EBPF, I realized that I could use the bpf program
> to categorize packets in ways packet-filter(7) does not provide.
>
> As a concrete example, I plan to build a util `rxtxmark` which could be
> passed something like `--mark-list 42,88`. This would be translated to a bpf
> program where the return code is the ordinality of the mark in the list.
>
> if (skb->mark == 42)
> return 1;
> if (skb->mark == 88)
> return 2;
> return 0;
>
> Packets enqueued to fd0 are simply ignored. Packets enqueued to the other fds
> are processed into pcaps and statistics.
>
> While I may build a util for tc_classid which does per-user-requested-classid
> pcaps and statistics like `rxtxmark` does for marks, I'm also interested in
> using tc_classid as a simple way to capture tx packets from a long running
> program on the fly.
>
> The program under inspection would simply be added to a net_cls cgroup which
> has a unique classid defined. A bpf program would be attached to map packets
> with that classid to fd1. While I can do this already by using iptables to
> translate the tc_classid to a mark, that complicates the implementation
> greatly since the firewall has to be touched (which is probably overreaching
> for a packet capture util and would most likely be left to the user to
> configure).
>
And only now do I discover netsniff-ng; a seriously cool set of utils! Thank
you for your efforts there Daniel!
I still plan to continue advancing my various PACKET_FANOUT utils and
eventually seeing how much, if any, of the common code would be of interest to
the libpcap maintainers. But very cool that a quick look at the netsniff-ng
help file shows that rxtxcpu et al could be accomplished with the right number
of concurrent invocations of netsniff-ng.
> > > I built a kernel (v5.1) with 4 instances of the following line removed
> > > from net/core/filter.c to test the behavior when the instructions pass
> > > verification.
> > >
> > > switch (off) {
> > > - case bpf_ctx_range(struct __sk_buff, tc_classid):
> > > ...
> > > return false;
> > >
> > > It appears skb->tc_classid is always zero within my bpf program, even
> > > when I verify by other means (e.g. netfilter) that the value is set
> > > non-zero.
> > >
> > > I gather that sk_buff proper sometimes (i.e. at some layers) has
> > > qdisc_skb_cb stored in skb->cb, but not always.
> > >
> > > I suspect that the tc_classid is available at l3 (and therefore to utils
> > > like netfilter, ip route, tc), but not at l2 (and not to AF_PACKET).
> >
> > From tc/BPF context you can use it; it's been long time, but I think back
> > then
> > we mapped it into cb[] so it can be used within the BPF context to pass skb
> > data
> > around e.g. between tail calls, and cls_bpf_classify() when in
> > direct-action mode
> > which likely everyone is/should-be using then maps that skb->tc_classid u16
> > cb[]
> > value to res->classid on program return which then in either
> > sch_handle_ingress()
> > or sch_handle_egress() is transferred into the skb->tc_index.
> >
>
> It sounds like just before the start of a BPF_PROG_TYPE_SCHED_CLS bpf program
> tc_classid id placed in skb->cb. The missing plumbing to support my use case
> is probably the same thing, but for BPF_PROG_TYPE_SOCKET_FILTER.
>
> I'll see about familiarizing myself with both as time permits and perhaps I
> can get tc_classid working for a BPF_PROG_TYPE_SOCKET_FILTER program; it
> certainly sounds like it's doable.
>
> > > Is it impractical to make skb->tc_classid available in this bpf context
> > > or is there just some plumbing which hasn't been connected yet?
> > >
> > > Is my suspicion that skb->cb no longer contains qdisc_skb_cb due to
> > > crossing a layer boundary well founded?
> > >
> > > I'm willing to look into hooking things together as time permits if it's
> > > a feasible task.
> > >
> > > It's trivial to have iptables match on tc_classid and set a mark which is
> > > available to bpf at l2, but I'd like to better understand this.
> > >
> > > Thanks,
> > > Matt C.
> > >
