On Wed, Jan 12, 2022 at 8:13 PM Toke Høiland-Jørgensen <toke@xxxxxxxxxx> wrote:
>
> [ adding the bpf list - please make sure to include that when sending
> BPF-related patches, not everyone in BPF land follows netdev ]
>
> "Jason A. Donenfeld" <Jason@xxxxxxxxx> writes:
>
> > BLAKE2s is faster and more secure. SHA-1 has been broken for a long time
> > now. This also removes quite a bit of code, and lets us potentially
> > remove sha1 from lib, which would further reduce vmlinux size.
>
> AFAIU, the BPF tag is just used as an opaque (i.e., arbitrary) unique
> identifier for BPF programs, without any guarantees of stability. Which
> means changing it should be fine; at most we'd confuse some operators
> who have memorised the tags of their BPF programs :)
>
> The only other concern I could see would be if it somehow locked us into
> that particular algorithm for other future use cases for computing
> hashes of BPF programs (say, signing if that ends up being the direction
> we go in). But obviously SHA1 would not be a good fit for that anyway,
> so the algorithm choice would have to be part of that discussion in any
> case.
>
> So all in all, I don't see any issues with making this change for BPF.
Somewhat related, if BPF is going to move from SHA to something, then
consider SipHash. Here are the numbers I regularly observe. They
remain relative the same on 64-bit platforms:
* SHA-1: 4.31 cpb using SSE2
* BLAKE2s: 4.84 cpb using SSE4.1
* BLAKE2b: 3.49 cpb using SSE4.1
* SipHash 2-4: 1.54 cpb using C/C++
* SipHash 4-8: 2.55 cpb using C/C++
If BPF is Ok with 64-bit tags, then SipHash 2-4 is probably what you
want on the wish list.
Jeff
