Re: BPF for hash calculation

On May 19, 2020, at 22:13, Ariel Levkovich [off-list ref] wrote:

Hi Daniel, Aleksei

I’m working on a feature to add support for datapath hash actions and matching via TC API.

One of the options we want to offer users there is to provide their own hash calculation function, in the form of a BPF program.
The program should accept struct __sk_buff and return a hash value.

After a little research and testing I noticed that some key parameters in struct __sk_buff are sertricted to BPF_PROG_TYPE_SK_SKB program types while I was planning to re-use the existing  BPF_PROG_TYPE_SCHED_ACT program type for the act_hash purpose.
The key parameters I’m referring to are fields like remote_ip4, local_ip4, src/dst_port (flow key fields) that are most likely going to be relevant for hash calculation on a packet.

So my question to you is basically what is the best option here? The way I see it the options are:
1. Remove restrictions on these fields for SCHED_ACT program type in kernel/bpf/verifier.c
2. Add new program type SCHED_ACT_HASH with permission to access these fields.
3. More of a question - is there another way to access the flow keys via struct __sk_buff?

Appreciate your advice and thanks in advance.

Best Regards,

Ariel Levkovich
Staff engineer, Mellanox SW

Forgot to CC relevant mailing list.
After further digging it seems there’s a pointer to a flow keys struct that is not restricted to SK_SKB prog types and should be our best option. The question remains whether we can reuse SCHED_ACT type or a new type for hash calculation should be added?