Re: [PATCH net-next 1/2] bpf: introduce bpf_skb_vlan_push/pop() helpers
flat view
From: Eric Dumazet <hidden>
Date: 2015-07-17 08:12:23
On Thu, 2015-07-16 at 19:58 -0700, Alexei Starovoitov wrote:
In order to let eBPF programs call skb_vlan_push/pop via helper functions
Why should eBPF program do such thing ? Are BPF users in the kernel expecting skb being changed, and are we sure they reload all cached values when/if needed ?
eBPF JITs need to recognize helpers that change skb->data, since skb->data and hlen are cached as part of JIT code generation. - arm64 JIT is using bpf_load_pointer() without caching, so it's ok as-is. - x64 JIT recognizes bpf_skb_vlan_push/pop() calls and re-caches skb->data/hlen after such calls (experiments showed that conditional re-caching is slower). - s390 JIT falls back to interpreter for now when bpf_skb_vlan_push() is present in the program (re-caching is tbd).
quoted hunk ↗ jump to hunk
+static u64 bpf_skb_vlan_push(u64 r1, u64 vlan_proto, u64 vlan_tci, u64 r4, u64 r5) +{ + struct sk_buff *skb = (struct sk_buff *) (long) r1; + + if (unlikely(vlan_proto != htons(ETH_P_8021Q) && + vlan_proto != htons(ETH_P_8021AD))) + vlan_proto = htons(ETH_P_8021Q);
This would raise sparse error, as vlan_proto is u64, and htons() __be16 make C=2 CF=-D__CHECK_ENDIAN__ net/core/filter.o