On 2/28/23 5:56 AM, Alexei Starovoitov wrote:
On Mon, Feb 27, 2023 at 5:57 PM Daniel Xu dxu@dxuuu.xyz wrote:
On Mon, Feb 27, 2023 at 03:03:38PM -0800, Alexei Starovoitov wrote:
On Mon, Feb 27, 2023 at 12:51:02PM -0700, Daniel Xu wrote:
=== Context ===
In the context of a middlebox, fragmented packets are tricky to handle. The full 5-tuple of a packet is often only available in the first fragment which makes enforcing consistent policy difficult. There are really only two stateless options, neither of which are very nice:
Enforce policy on first fragment and accept all subsequent fragments. This works but may let in certain attacks or allow data exfiltration.
Enforce policy on first fragment and drop all subsequent fragments. This does not really work b/c some protocols may rely on fragmentation. For example, DNS may rely on oversized UDP packets for large responses.
So stateful tracking is the only sane option. RFC 8900 [0] calls this out as well in section 6.3:
Middleboxes [...] should process IP fragments in a manner that is consistent with [RFC0791] and [RFC8200]. In many cases, middleboxes must maintain state in order to achieve this goal.=== BPF related bits ===
However, when policy is enforced through BPF, the prog is run before the kernel reassembles fragmented packets. This leaves BPF developers in a awkward place: implement reassembly (possibly poorly) or use a stateless method as described above.
Fortunately, the kernel has robust support for fragmented IP packets. This patchset wraps the existing defragmentation facilities in kfuncs so that BPF progs running on middleboxes can reassemble fragmented packets before applying policy.
=== Patchset details ===
This patchset is (hopefully) relatively straightforward from BPF perspective. One thing I'd like to call out is the skb_copy()ing of the prog skb. I did this to maintain the invariant that the ctx remains valid after prog has run. This is relevant b/c ip_defrag() and ip_check_defrag() may consume the skb if the skb is a fragment.
Instead of doing all that with extra skb copy can you hook bpf prog after the networking stack already handled ip defrag? What kind of middle box are you doing? Why does it have to run at TC layer?
Unless I'm missing something, the only other relevant hooks would be socket hooks, right?
Unfortunately I don't think my use case can do that. We are running the kernel as a router, so no sockets are involved.
Are you using bpf_fib_lookup and populating kernel routing table and doing everything on your own including neigh ?
Have you considered to skb redirect to another netdev that does ip defrag? Like macvlan does it under some conditions. This can be generalized.
Recently Florian proposed to allow calling bpf progs from all existing netfilter hooks. You can pretend to local deliver and hook in NF_INET_LOCAL_IN ? I feel it would be so much cleaner if stack does ip_defrag normally. The general issue of skb ownership between bpf prog and defrag logic isn't really solved with skb_copy. It's still an issue.
I do like this series and we would also use it for Cilium case, so +1 on the tc BPF integration. Today we have in Cilium what Ed [0] hinted in his earlier mail where we extract information from first fragment and store the meta data in a BPF map for subsequent packets based on ipid [1], but limitations apply e.g. service load-balancing won't work. Redirecting to a different device or moving higher up the stack is cumbersome since we then need to go and recirculate back into tc BPF layer where all the business logic is located and handling the regular (non-fragmented) path, too. Wrt skb ownership, can you elaborate what is a concrete issue exactly? Anything that comes to mind with this approach that could crash the kernel?
[0] https://lore.kernel.org/bpf/cf49a091-9b14-05b8-6a79-00e56f3019e1@gmail.com/ [1] https://github.com/cilium/cilium/pull/10264