I would be interested to participate in the BPF track. The following topics would be especially interesting to cover, - Having spent the last several months writing increasingly complex BPF code we have managed to stumble into a handful of instances where the verifier, compiler and user appear to be fighting each other to generate code that is both enjoyable C to the user, able to be compiled and optimized by LLVM without throwing errors and warnings, and finally is verifiable from the kernel side. At times the only workaround we have found so far is to create hand written asm blocks of code. Fortunately, LLVM supports asm goto now! Some of the work to improve this is already in flight and should be in BPF trees by the workshop but some items we suspect will still be open by April. I think it would be an interesting session to collect a set of use cases (C code samples) that can not be compiled and loaded to understand where the pain points in LLVM and kernel verifier are today so we can improve them going forward. Of course, we could compile a list of just our examples but ideally (assuming others find the topic interesting) we could try to collect a bigger set of examples from the community. - A related item is lack of alloc/realloc in BPF keeps causing rather odd hacks in our use cases. For example, one common item that continues to pop up is we over allocate map entries to account for worst case scenarios. Even when smaller entries would work the majority of the time. Or when doing probe_read_str() parsing we have no way to realloc buffers if the string is longer than expected. Here we would like to propose a scheme for supporting malloc/realloc/free APIs from BPF programs. - Finally, changing topics a bit Cilium has implemented an encryption layer using the stacks IPsec layer. However, this results in passing packets into the stack and out of the stack complicating both the datapath and control plane. Additionally, wireguard has recently been added. A native solution for encryption in BPF for both IPsec and wireguard would greatly simplify integrating encryption with BPF including at the XDP layer. Also it looks like a session on sockmap and related socket based BPF hooks would be useful based on other proposals. I would like to attend to provide input here as well. Thanks, John
