Hi Martin, > > - bpf_skops_tx_timestamping(sk, skb, op, 2, args); > > + if (sk_is_tcp(sk)) > > + args[2] = skb_shinfo(skb)->tskey; > > Instead of only passing one info "skb_shinfo(skb)->tskey" of a skb, pass the > whole skb ptr to the bpf prog. Take a look at bpf_skops_init_skb. Lets start > with end_offset = 0 for now so that the bpf prog won't use it to read the > skb->data. It can be revisited later. > > bpf_skops_init_skb(&sock_ops, skb, 0); > > The bpf prog can use bpf_cast_to_kern_ctx() and bpf_core_cast() to get to the > skb_shinfo(skb). Take a look at the md_skb example in type_cast.c. In recent days, I've been working on this part. It turns out to be infeasible to pass "struct __sk_buff *skb" as the second parameter in skops_sockopt() in patch [11/11]. I cannot find a way to acquire the skb itself, sorry for that :( IIUC, there are three approaches to fetch the tskey: 1. Like what I wrote in this patchset, passing the tskey to bpf prog through calling __cgroup_bpf_run_filter_sock_ops() is simple and enough. 2. Considering future usability, I feel I can add skb_head, skb_end fields in sock_ops_convert_ctx_access(). 3. Only adding a new field tskey like skb_hwtstamp in sock_ops_convert_ctx_access() If there is something wrong, please correct me. Thanks! patch[11/11]: https://lore.kernel.org/all/20241207173803.90744-12-kerneljasonxing@xxxxxxxxx/
