On Tue, 19 Nov 2019 at 16:44, Jason A. Donenfeld <Jason@xxxxxxxxx> wrote: > > On Tue, Nov 19, 2019 at 4:34 PM Ard Biesheuvel > <ard.biesheuvel@xxxxxxxxxx> wrote: > > > - Zinc's generic C implementation of poly1305, which is faster and has > > > separate implementations for u64 and u128. > > I assume your AndyP comment below didn't apply to this top item here. > This one should be fairly uncontroversial in your opinion, right? > Yes. > > > - x86_64 ChaCha20 from Zinc. Will be fun to discuss with Martin and Andy. > > > - x86_64 Poly1305 from Zinc. > > > > As I pointed out in the private discussions we had, there are two > > aspects two AndyP's benchmarking that don't carry over 100% to the > > Linux kernel: > > - Every microarchitecture is given equal weight, regardless of the > > likelihood that the code will actually run on it. This makes some > > sense for OpenSSL, I guess, but not for the kernel. > > - Benchmarks are typically based on the performance of the core > > cryptographics transformation rather than a representative workload. > > This is especially relevant for network use cases, where packet sizes > > are not necessarily fixed and usually not a multiple of the block size > > (as opposed to disk encryption, where every single call is the same > > size and a power of 2) > > > > So for future changes, could we please include performance numbers > > based on realistic workloads? > > Yea I share your concerns here. From preliminary results, I think the > Poly1305 code will be globally better, and I don't think we'll need an > abundance of discussion about it. > Good. > The ChaCha case is more interesting. I'll submit this with lots of > packet-sized microbenchmarks, as well as on-the-wire WireGuard > testing. Eric - I'm guessing you don't care too much about Adamantium > performance on x86 where people are probably better off with AES-XTS, > right? Are there other specific real world cases we care about? IPsec > is another one, but those concerns, packet-size wise, are more or less > the same as for WireGuard. But anyway, we can cross this bridge when > we come to it. > As long as we can show that WireGuard really benefits without regressing other users disproportionately, I think we're fine. There are not /that/ many users to begin with ... > > > > - WireGuard! Hurrah! > > > > > I'm a bit surprised that this only appears at the end of your list :-) > > Haha, "last but not least" :) > > > > > > If you have any feedback on how you'd like this prioritized, please > > > pipe up. For example Dave - would you like WireGuard *now* or sometime > > > later? I can probably get that cooking this week, though I do have > > > some testing and fuzzing of it to do on top of the patches that just > > > landed in cryptodev. > > > > > > > We're at -rc8, and wireguard itself will not go via the crypto tree so > > you should wait until after the merge window, rebase it onto -rc1 and > > repost it. > > Thanks, yea, that makes sense. Netdev also has its own merge window > schedule that I should aim to meet. I guess, based on this if I'm > understanding correctly, we're looking at WireGuard for 5.5? > No, v5.6. The merge window for v5.5 will open this Sunday, so you'll need to rebase on v5.5-rc1 once it is released, which will [hopefully] have all the crypto pieces you need. Note that this applies equally to the other changes: we can queue performance tweaks in the crypto tree in parallel, and they will all hit v5.6 at the same time.
